APBRmetrics

When evaluating a player through the lens of Adjusted Plus Minus, how valuable do you find minutes to be?

For example. is a player with a RAPM of 1.7 averaging 32 minutes per game a better player than another player with a RAPM of 2.2 but only 18 minutes per game?

Is Cory Joseph's RPM of .88 less meaningful because he only played 18 minutes per game? It it less meaningful in light of the fact that the Spurs didn't play him in the playoffs?

Conceptually, I think there has to be some sort of trade-off between RAPM values and minutes played, but I have no idea what the number should be.

If you had the estimated errors or even estimates of the estimated errors, you 'd be in better position to answer the example question. We don't have the estimated errors but using estimates of the estimated errors, one could construct a decent tradeoff equation / matrix for which player is more likely better. It is just a more likely better based on probability call though, not a definitive he is better statement though. Looking at multi-year RAPM and each players entire RAPM / RPM should be part of the plan, instead of staying within one year's estimate, imo. Look at the box score metrics too. And the "context" of the player's performance.

Actually on a related note, I had a question on teammate RAPM. Would it be expected that two separate players, when joining a team, should remain at the same RAPM? For example a +4 and a +4 on separate teams, if they were to play together would they in a perfect world be +8 or less? And if less, is that showing anything (ie that they aren't a "fit" together)?

bondom343 wrote:Actually on a related note, I had a question on teammate RAPM. Would it be expected that two separate players, when joining a team, should remain at the same RAPM? For example a +4 and a +4 on separate teams, if they were to play together would they in a perfect world be +8 or less? And if less, is that showing anything (ie that they aren't a "fit" together)?

For one thing, they might perform worse than +8 because of the 'effect of leading' - unless the rest of the team performs at -8 total, in which case it'd balance out. Although I know that doesn't exactly answer your question

So far, most if not all the research I've done points at the fact that players don't have much influence on each other's impact. I find that hard to believe, myself, so there's more work to be done

On a sidenote, this is not exactly a problem that, if it exists, exists only with RAPM. Two players with a PER of 27 joining the same team might not perform at a PER of 27 from then on.

Common sense would tell us to look at the player's skill set & positions. Chris Paul and Ty Lawson probably isn't the greatest fit, but CP3 can definitely work with almost any non-high-usage player

Thanks JE, appreciated!

Impact on teammates or new teammates probably should be done at Factor level. The degree of teammate might be greater on defense than offense. On offense the impact might be more on efg% and turnovers than offensive rebounds and getting to the line.

There has been plenty of debate of value of RAPM / RPM but the value of it depends a lot on the intensity of the analysis using it after you have it. It is as much or more stepping stone than final product.

I had been thinking about this issue somewhat recently and was trying to find a particular related post, but with no success. Wasn't there a recent discussion (within the past year?) where very interesting evidence was presented relating plus minus performance to the duration of on-court stints? Might anyone here have the link at hand to share?

As I recall, the variation presented was very large (i.e. competitively meaningful). Wasn't it something like +/- 2 relative to average performance, with almost linear gains with on-court time to about 6 or 8 minutes?

Stipulating something like this, it should matter a lot in terms of estimating the "true" xRAPM, with players having disproportionate number of short stints on the court appearing worse than similarly able players playing a greater proportion of optimal length.

This one? viewtopic.php?f=2&t=8614&hilit=stint+rapm

So your suggesting further adjusted adjusted plus minus. Maybe should for stint distribution and perhaps other things.

Look at box score, sportsvu, rpm, rpm factors, further processed rpm. As much as teams seem to spend time and money to build data products / applications and data analytics, there should be a ton of time spent on contemplation on those things, comparison, manipulation, combination, revision, characterization, explanation, salesmanship, etc.

Yes, Crow, that's it. Thanks! So my memory of the dimension of the effect was more or less correct; the range was about 3 (not 4) with stints less than 2 minutes duration having a combined O and D effect of -2.1 rising to a plateau of +0.7 between 6 and 10 minutes duration.

And it seems to me that this "warmup" effect is kind of a big deal that one would want to incorporate in xRAPM estimates. Not as large (as I recall) as Jeremias' scoring margin "effort" effect, but surely (?) relevant for better sorting out player abilities. And perhaps these effects interact in interesting ways? (Might this effect already be incorporated in the mysteriously-generated RPM?)

Even if the adjusted plus minus isn't adjusted, it would be interesting to see how the proportion of total time that is first 2 minutes of stint for players and how it varies by team. I guess I already said in the referenced thread that JE's data is a surface argument for minimizing the number of lineup substitutions and probably the number of lineups used too. Know your minute allocations and achieve them efficiently without so much micro-tactical lineup variation / bloat in lineups used / in&out first 2 minute efficiency loss.

Crow wrote:Impact on teammates or new teammates probably should be done at Factor level. The degree of teammate might be greater on defense than offense. On offense the impact might be more on efg% and turnovers than offensive rebounds and getting to the line.

There has been plenty of debate of value of RAPM / RPM but the value of it depends a lot on the intensity of the analysis using it after you have it. It is as much or more stepping stone than final product.

Yeah, it was from a separate debate I was having, mainly on Westbrook and Durant. The argument made to me was that Durant's RAPM was highest in the year he played without Westbrook, and Westbrook's in the year without Durant, so they aren't a good fit with each other. I said I didn't think they were really related and it happened to be due to the situations at hand.

If the efg% factor of offensive RAPM / RPM is correlated with usage and "creation" plays(and I would guess that they are) then if KD or Russ has to play without the other, their usage and opportunity for creation is likely to be higher and so would be the personal rating- if they are capable of taking a bigger load alone than what they have together.

There are usage - offensive efficiency curves. If one calculated RAPM / RPM for usage bands and creation bands (essentially splitting the player into different versions, probably using multiple seasons of data to offset sample shrinking), one could have curves plotting that behavior vs estimated impact on the RAPM / RPM. It could help with seeing optimal role size.

Going back to my original post:
Thinking about RAPM with a Bayesian outlook -- RAPM is not an evaluation of talent or skill, but just the measure of impact a player has while on the court given the context of the game/other 9 players on the court; less minutes per game (not less minutes on the season though) means less contexts that a player will be exposed to while evaluating their impact. As coaches have a preference for putting players in a situation where they will excel rather than putting players in a position to fail, this implies that given the same RAPM value, the player with higher minutes per game should be regarded as superior.

NOTE - this has nothing to do with minutes per season and an increased sample size (which is an important, but completely separate issue).

To my understanding, the "fewer contexts" argument, as described, speaks only to the precision of the estimates not their accuracy. However, the "fewer contexts" being correlated with shorter on-court stints (which I'm sure it is) I think implies an opposite empirical effect than the one you are supposing.

If there is indeed an substantial "warm-up" effect, and fewer on-court minutes imply shorter stints, then such players are being unduly penalized for playing "cold".

Also, I don't think the argument for "fewer context" players having an advantage owing to their being disproportionately put in advantageous situations is correct. At best it is partial, as opposing coaches have the opposite motivations.

I think a fundamental issue here is that you can't answer all questions well with a single statistical measure.

So far, most if not all the research I've done points at the fact that players don't have much influence on each other's impact. I find that hard to believe, myself, so there's more work to be done.

I'm assuming that you mean "the difference in the team effects of various NBA players seems to be small". Playing 4-on-5 is probably not going to go so well for the 4-man team.

Common sense would tell us to look at the player's skill set & positions. Chris Paul and Ty Lawson probably isn't the greatest fit, but CP3 can definitely work with almost any non-high-usage player

Yeah, it's pretty clear that position relative to the ball and position relative to the hoop are big factors in how much impact a player can have on the game - especially as measured by box scores. If nothing else, inside players rebound more, and players have to have the ball to shoot. Ostensibly something like SportsVU would help control for opportunity. A more subtle aspect is separating the impact of the players' role from the impact of the players' ability.