+/- Questions

[HoopDon]

Just had a few question in relation to a post I'm trying to write.

What is the minimum number of games RAPM needs for solid results? What about APM? Raw +/-? Any links to articles/research involving these questions would be great.

http://hoopdon.weebly.com/

[Mike G]

The cynic in me wants to reply that the number of games needed is such that by the time you have them, the players are no longer what they were.

[J.E.]

What, exactly, do you classify as a 'solid result'?

APM needs at least two years, and maybe more. If you have less you might as well roll a dice for every player

[Crow]

A plus minus estimate for a starter for a full season is about as solid is you can get from the metric. For half the season it will be less but I think the average error will not be fully twice as high. Ratings for a 20 minute per game sub will be more accurate than for a 12 or 8 minute sub. If you have played 2000 minutes in a season the estimate is likely to be within 1 - 2 of true impact at least
2/3rds of the time. I think. Prior informed APM is arguably more accurate because it based on more minutes but is debatable as Mike's comments suggests. Some would frame their reply to this in terms of possessions but I find it easier to use minutes as a gauge.

[HoopDon]

So two years for APM, and raw +/- pretty much never?

I recall reading somewhere that 50 games or so was the minimum "acceptable" sample size for RAPM. Any truth to this?

http://hoopdon.weebly.com/

[J.E.]

So two years for APM, and raw +/- pretty much never?

I recall reading somewhere that 50 games or so was the minimum "acceptable" sample size for RAPM. Any truth to this?

http://hoopdon.weebly.com/

Let me say that APM and raw +/- should never be used when RAPM exists. If you don't have the means to compute (R)APM you could probably bake raw +/- into a somewhat decent metric but you'd have to make small changes to it, i.e. dividing by 5 (because 5 players are on the court for one team) and taking NET into account, while also regressing to the mean (less player minutes -> more regression to the mean).

50 games is a rather arbitrary cutoff. Because of RAPM's built in regression to the mean, any number of games will give you player ratings that are better than rolling a dice (although not a lot better if you only feed it a handful of games. APM, in contrast, is worse than rolling dice if you don't feed it a minimum number of games).
RAPM standard errors go down almost linearly in relation to minutes between 0 and 5000 minutes. Then things starts to flatten

[Mike G]

It kind of changes the perspective on what constitutes "small sample size". In 50+ games, or 1500 minutes, a player's rebounding, FT%, etc are pretty well established; but not his overall effectiveness.

At the other end of the spectrum, a head coach in the playoffs has to discern within minutes what seems to be working on the court.

[DSMok1]

So two years for APM, and raw +/- pretty much never?

I recall reading somewhere that 50 games or so was the minimum "acceptable" sample size for RAPM. Any truth to this?

http://hoopdon.weebly.com/

Let me say that APM and raw +/- should never be used when RAPM exists. If you don't have the means to compute (R)APM you could probably bake raw +/- into a somewhat decent metric but you'd have to make small changes to it, i.e. dividing by 5 (because 5 players are on the court for one team) and taking NET into account, while also regressing to the mean (less player minutes -> more regression to the mean).

50 games is a rather arbitrary cutoff. Because of RAPM's built in regression to the mean, any number of games will give you player ratings that are better than rolling a dice (although not a lot better if you only feed it a handful of games. APM, in contrast, is worse than rolling dice if you don't feed it a minimum number of games).
RAPM standard errors go down almost linearly in relation to minutes between 0 and 5000 minutes. Then things starts to flatten

A little bit off topic, but: at what point, in your opinion/experience, will RAPM (without box score information) be better (more accurate) than an SPM metric? How many games/years would it take? Would it happen at all?