APBRmetrics

[tenkev]

Its been low all season. At the end of the season it was -0.06 and including the playoffs its at 1.67. Peja's +/-, on the other hand, is at 11.35 right now. Obviously Chris Paul is better than his stat and Peja is not nearly as good as his; but, could it be possible that Peja is actually the more important player on that team?

Could it be that Paul's inability to contest shots or rebound the ball are such a liability that he is not as valuable as PER, BoP stats and all other box score based stats make him appear to be?

[thref23]

Where do you get your info from? Per 82games.com Paul's offensive +/- was +15.4, and total +/- +9.3. Peja's was +12.8 offensively, +9.8 overall, but thats partially because opponents are more likely to rest their best players with Paul out of the game (my best guess). In the postseason, his offensive impact is +17.4, overall +9.9

Now, his playoff +/- is nothing compared to Tyson Chandler's +42.8 - but this is utilizing a small sample size.

Unless you somehow confused Chris Paul's info with MoPete's, you must be going by adjusted +/-. I'm not sure how that works out considering that Peja had only played slightly more than 200 minutes this season without Chris Paul in the game, and indications would be that he had a negative +/- in those 200+ minutes. Chris Paul played about 500+ minutes without Peja, and had a positive +/- in those minutes.

http://82games.com/0708/0708NOHP.HTM

My guess is, in the 700+ minutes that the two weren't paired together, the adjusted +/- formula is at fault by either overweighting and/or miscalculating the quality of supposed competition on the floor and/or is simply negatively skewed by certain small sample sizes. Peja also benefits here, in general, since the Hornets lack solid wing depth behind him.

Peja's perimeter presence is what makes the Hornets' offense officially tick, as it would not work nearly so well if it had to go through MoPete and/or say Rasual Butler or Julian Wright all the time to stretch the opposition out. But adjusted +/- is simply off-base here.

[Mike G]

According to this page:
http://basketballvalue.com/teamplayers.php?year=2007-2008&team=NOH

... all of the Hornets' guards are in the red (this is for regular season only), while all their frontcourters (except Hilton Armstrong) are in the green (positive). Peja several times better than anyone else, at +13.32 .

How exactly can this be? Have they gone stretches without a guard or two, and done very well in those minutes? Starters and their backups all look bad, relative to when they are out of the game. Weird.
_________________
`
36% of all statistics are wrong

[Kevin Pelton]

[Neil Paine]

I have to think it's some combination of a fluke and a misattribution of Paul's contributions to Stojakovic. In 2005-06, Paul's APM was +10.22, and in 2006-07 it was +4.14, both of which track fairly well with his boxscore stats. So I find it hard to believe that in this season, when his boxscore stats are through the roof, he suddenly has a negative impact on the team, while his teammate Stojakovic (+1.85 in 2007, +0.34 in 2006, -0.33 in 2005) is suddenly the 2nd-best player in the league.

[bchaikin]

It's worth noting that Paul's on-off disparity is the worst of all Hornets players (starters or subs), so it cannot be a mere artifact of his being on the court with other poor defensive players

the truehoop bullet goes on to quote the author:

...the best measure of a player's overall impact on the game's bottom line is his adjusted plus-minus rating... Put simply: his superb offensive contributions have been largely offset by his apparent defensive liabilities.

nonsense...

paul's offensive numbers in 07-08 were very good. his points scored per zero point team possession personally responsible for (through missed FGAs or FTAs rebounded by the defense or by committing turnovers) was among the best of all PGs - i.e. he was very efficient on offense - and he was one of the top scoring PGs in the league...

he also in 07-08 got 217 steals in 3006 minutes. the league average for a player playing 3006 minutes this year was 91 steals per 3006 minutes. so that's 126 defensive stops that paul got over and above what just the league average player got playing the same minutes. had for example steve nash played 3006 minutes he'd have gotten somewhere around 57-58 steals, so that'd be 159-160 more defensive stops than what nash got simply through more steals alone...

plus paul was a slightly better defensive rebounder than just the league average PG, on a very slow paced team (i.e. he didn't get more defensive rebounds simply because the hornets played at a fast pace)...

so for paul to have been a poor defensive player overall (enough to offset his very good offensive contributions) yet with a high number of defensive stops through steals alone, he'd had to have given up an outrageously high FG% to the opposing players he guarded while on defense, or allowed an outrageously high amount of points scored, to be considered a worse defensive PG than just the league average PG, let alone the majority of the league's PGs. i don't see any evidence for this...

the poor adjusted plus-minus is at least something of a fluke.

agreed - it simply isn't accounting for something if its coming to the conclusion that paul was a poor defensive player overall in 07-08...

[Mountain]

It would indeed help for the experts on adjusted +/- to provide more commentary.

Adding to the data observations in this thread I will note that Paul with the other 4 starters have a fine raw and adjusted 5 man lineup (the later according to countthebasket). This one lineup represents 40% of his time and his player pair time is very high with each of the 4.

I do also see the 4 other starters plus Pargo instead of Paul has a near identical raw +/-, but this was in a small sample of 40 minutes. And the other 4 with Bobby Jackson, in a whopping 6 minutes, was actually twice as good. I don't know how big an affect this has on Paul's adjusted +/- outcome.

I assume this is a more difficult than usual case for the method, perhaps akin to the Detroit case that D Lewin described in his article.


The low adjusted rating is perplexing. He is roughly tied with Peja for best raw +/- and has positive player pairs with everyone and consistently strong scores for every partner except Armstrong.

The raw on/off data makes it seem that most or all of the problem is in FG% allowed. He gave up a higher FG% to his man than Pargo and Jackson and at team level too. He gives up a bit worse opponent PER than both. Adjusted 4 factors for Paul, if possible and done, might reveal more specifically and confidently the size of these negatives and his positives individual and team.

[Harold Almonte]

The +/- of players like Paul and Kobe, who play above 75% of available minutes, should be taken with a double grain of salt. The quality of the replacement time could be having more weight than the quality of the replacement player.

Kobe has also a negative defensive influence according to that. One could argument that Paul is average defender, but not Kobe.

[Harold Almonte]

Also, somebody said they are defending the shot (opp.FG% holding) a 3 or 4% below their positions's average. I don't know the reliance of these averages, but if somebody is tracking and scorekeeping defensive attempts, play by play, then all defense metrics's problems are already solved.

[Mountain]

Peja #2 in league on adjusted +/-. ...

[Mountain]

The size of the standard error makes it hard to put that much weight on these numbers, as has been cautioned properly before.

I recently computed that 69% of all NBA minutes are played by lineups in use less than 100 minutes. Knowing now that there is so high a dependence on small sample lineups as input for adjusted +/- ratings makes me much more cautious about using pure adjusted +/- scores.

Overall +/- including statistical +/- probably is better.

My impression is that player performance in different 5 man lineups performance is pretty volatile and team performance is not that accurately estimated from the sum of adjusted average player impact scores. But I'll look into that more and would be interested in seeing any presentations about that.

Coaches have decent data on very very few lineups. The rest require use of the individual stats / evaluations and feel or guessing about the interactions and the performance of the whole.

[Mike G]

I still don't see how a team that averaged 5.2 Pts/48 more than their opponents could carry a squadron of guards whose +/- were ALL negative. Doesn't there have to be an error in some calculation?
_________________
`
36% of all statistics are wrong

[Mountain]

A quick look at the player and lineup adjusted datasets together:

For the most used lineup in NBA Paul-Peterson-Stojakovic-West-Chandler
the sum of the basketballvalue reported player adjusted +/-'s and the countthebasket simple adjusted for the lineup are only modestly different (2).

But it is, as expected, more volatile with smaller minute lineups.

Replace Paul with Pargo for season and the sum of the players adjusted suggests a decline by about 4.5 but the adjusted for the lineup declined by 19. The +/- for the actual usage of the lineup didn't change at all despite the fact they faced lineups with very good ratings. But in a small sample of course almost anything can happen.

However replace Paul with Jackson and the sum of the adjusted for the players and the adjusted for the lineup and the raw +/- change about equally. But this may be more coincidence than accurate transwalk.

Replace Chandler with Armstrong and the player adjusted suggests a reduction in average adjusted player impact of 11 pts but the Armstrong lineup shows 27 pts weaker. And the actual data shows a change of 37 (before adjustment for opponent quality).

Replace Peterson with R Butler and the player numbers suggest a change of 9 but the adjusted lineup shows 14. The actual changed by 10. That's pretty tight.

Replace Peterson with Pargo the player numbers suggest a slippage of 2 but the adjusted lineup shows 7. But the actual improved by 4.

The adjusted averages of the two methods try to get at underlying strength but do bounce around differently (subject to the different sample sizes and the unique synergy of a lineup, but how much is each ?). They are only rough guides on what to expect from the next sample of a particular lineup of course and lots of variance should be expected.

In the hands of a statistical expert I guess the information about the population and the sample can be balanced and worked to make judgments that might have a better chance of being right than the less informed.

However how much of prospective lineup management is based on data guidance and how much do coaches - despite whatever the plan was, after the game begins- mostly go to match-up predictions using the sum of their coaching knowledge / experience / instincts and retrospective coaching based on the results of the immediate past minutes? Who among coaches in the NBA is a notably disciplined follower of statistical guidance on player team impacts and lineup strength? How many GMs grade the coaching lineup management results vs the stat guidance? I'd love to hear from the consultants but ...

Things might not change much unless teams become a good deal more disciplined at lineup usage to increase the minutes of best prospect lineups and in turn improve the reliability of that 5 man data and then use the data to improve the lineup set. Can folks really provide much lineup guidance off adjusted scores for the most part of very small sample size lineups?

The guidance from that and the individual adjusted data may concur or may conflict as shown above. At this point I'd guess you'd generally favor the larger sample size individual score? Unless you were a big advocate that "basketball is a team game".

I have suggested previously that 3 and 4 man sets provide more minutes together and could perhaps increase knowledge of the performance of "cores" adjusting additionally around the other players involved. Time-consuming stuff. And a step away from unique 5 man lineups. But if you are looking at all the other adjusted information with assumptions and limitations, maybe this would be a decent additional step?

[Kevin Pelton]

Someone can correct me if I'm wrong, but I recall Dan Rosenbaum explaining it is a misconception (which we almost all had in the beginning) that the limited number of minutes for each lineup is responsible for the noisy nature of adjusted plus-minus ratings. The issue is that teams don't use more lineups and give more data points involving each player. I believe that teams that make midseason trades tend to have smaller standard errors.

[Mountain]

This concern about short minute samples is hard for me to shake off completely at this moment but if anyone can add further confirmation of this or add more on the topic, great.

Not a lot of credence goes to under 100 minute unique context samples in any other basketball analysis for players or lineups I am aware of. Players do not get rated on how they performed in a home and away against one team from another conference. And really these same own lineups are the sum of facing a set of different opposing lineups so it is really a bundle of tiny things, only half the same and the player under study is only 1 of 10 actors on the court contributing to the +/-.

Still I can understand that the combination of many many such samples would improve the statistical confidence level. I can also see that more lineups could help refine a player's impact but big minute players often have 150 - 200+ lineups which seems pretty good.

The standard errors are pretty large regardless of dominant reason for the noise. If the noise is "more" because insufficient lineups, ok. But small sample minutes are still not as helpful as larger ones would be.

It seems unlikely that the number of different lineups that will grow dramatically from the current pretty high (in my mind) levels so while that may be a more powerful theoretical help, I question the practicality of using regular season minutes for that, though preseason can be used for that, and blowouts and for some teams (depending on their rank and stage of development and goals) I guess some normal regular season time could be devoted to this data gathering too (possible discovery / possible price to pay). Rotations are already highly fractured with a pretty high % of possible lineups at least tried minimally so you'd have dig thru the bench to get to 4 or 5th guys at specific positions to increase the number of lineups tried for main guys or total lineups much beyond what we already see typically.

I still find it far easier to advocate expanding minutes of usage of the best prospect lineups from both a improving the reliability of the data for analysis standpoint and a trying to win games now standpoint.

Putting this technical background issue aside you still have the topic of whether coaches are reviewing adjusted 5 man lineup scores and giving much (or enough) weight to what the numbers say in how they run the rotation. Right now they probably can't as is beyond a few top used lineups. But they could understand a couple of handfuls more lineups better if they built up the sample sizes of the main lineup choices to get better reads on their effectiveness.

Chris Paul's and Peja's adjusted average +/-s over all minutes should be estimated and considered. A coach probably wants them on the court a lot in large part because of Peja's and despite Paul's.

The game is played with 5 man lineups and more minutes in a specific lineup / context would seem to be pound for pound the best guide on how player impact adds up and how the team would perform in the future in that specific lineup.