Berri Changes Value of Defensive Rebounds in WP

[EvanZ]

Man, took long enough. It's a shame it didn't happen years ago.

http://wagesofwins.com/2011/12/11/wins- ... -stronger/

The “proof” some people constructed focused on the value of rebounds. Because Wins Produced argues that non-scorers who grab many rebounds are quite valuable – and conventional wisdom argues that non-scorers are simply not that valuable – it must be the case that Wins Produced overvalues rebounds.

A variety of arguments have been offered in response to this critique. Many of these were detailed in the Frequently Asked Question page. For example, it was noted that

although diminishing returns – as detailed in Stumbling on Wins — certainly exists for defensive rebounds (but not for offensive rebounds), the size of the effect is “small”.
to illustrate, when the impact of diminishing returns with respect to defensive rebounds is accounted for, the ranking of the players doesn’t seem to change much (a point made on the FAQ page).
As a consequence, Wins Produced has historically ignored this issue.

As the lockout dragged on and on, though, I began to think that maybe it might be better to just incorporate this effect into the measure we post at the WoW Journal. After all, the effect has been measured and it can be included. So why not just make the adjustment and therefore remove the argument “Wins Produced overvalues rebounds” from the discussion.

So that is what I have done. The specific calculations – which are somewhat more complicated than what was posted before — are detailed at the new “Calculating Wins Produced” website.

Here's his rationale:

http://wagesofwins.com/wins-produced/ho ... -produced/

More specifically, for the above specification we see that each one-unit increase the teammate’s defensive rebounds reduces a player’s defensiverebounds per-minute by 0.504. In other words, some of a player’s defensive rebounds are taken from the player’s teammates. In contrast, on the offensive glass the teammate’s offensive rebounds per-minute do not have a statistically significant effect, suggesting that a player’s offensive rebounds do not come from the player’s teammates. These results make sense when we consider how teams go for rebounds on each end of the court. On the defensive end one suspects teams will send more players to the boards then they need because the team cannot begin its offensive possession if it doesn’t have the ball. Consequently, as already noted, one suspects that some defensive rebounds are just taken from a player’s teammates. On the offensive end, though, it’s a different story. Teams have to have some players back to play defense. So the offensive glass sees little competition among teammates, and therefore, offensive rebounds are primarily taken from a player’s opponent.

[DSMok1]

Top 10 before:
Kevin Love
Dwight Howard
Chris Paul
LeBron James
Dwyane Wade
Landry Fields
Pau Gasol
Kris Humphries
Zach Randolph
Kevin Garnett

Top 10 after:
Dwight Howard
Chris Paul
Kevin Love
LeBron James
Dwyane Wade
Pau Gasol
Steve Nash
Landry Fields
Rajon Rondo
Ray Allen


That looks better, though I think the position adjustment as relates to rebounding is still doing odd things (hence Landry Fields at 8 and Kris Humphries quite high).

[Mike G]

It's only half as bad as it was before?

[DSMok1]

I guess he's still not dealing with replacement level on rebounding?

[mystic]

I guess he's still not dealing with replacement level on rebounding?

No, he hasn't solved the real issue at all. The problem was never defensive rebounding alone. An important problem with his metric comes from the fact that his formula thinks a player can score points without using a possession. And while in reality the correlation coefficient between scoring margin and offensive rebounding over the last 10 years is around -0.2, Berri ends up with 0.2 as the correlation coefficient between offensive rebounding and WP48. When a player grabs an offensive rebound he used one possession less for his scoring according to Berri's formula used to calculate WP48. On a team basis offensive rebounding just keeps the possession, in Berri's world an offensive rebound is aquiring a possession.

The very basis of his approach is wrong. He is using formulas working on a team basis, but aren't useful for individuell players at all.

[xkonk]

I guess he's still not dealing with replacement level on rebounding?

No, he hasn't solved the real issue at all. The problem was never defensive rebounding alone. An important problem with his metric comes from the fact that his formula thinks a player can score points without using a possession. And while in reality the correlation coefficient between scoring margin and offensive rebounding over the last 10 years is around -0.2, Berri ends up with 0.2 as the correlation coefficient between offensive rebounding and WP48. When a player grabs an offensive rebound he used one possession less for his scoring according to Berri's formula used to calculate WP48. On a team basis offensive rebounding just keeps the possession, in Berri's world an offensive rebound is aquiring a possession.

The very basis of his approach is wrong. He is using formulas working on a team basis, but aren't useful for individuell players at all.

Don't all of the box score measures start with the team level and transfer down to individual players? At some level they must, since the value of a possession is calculated at the team level. And I have yet to see a metric that considers offensive rebounding to be a negative.

Also, I'm pretty sure that assists, for example, 'score' points without using a possession. All box score metrics assume that actions besides literally scoring are worth something, regardless of if they use a possession or not.

And if you knew how WP48 was calculated, you would know that offensive rebounds are subtracted off from possessions. They don't count as acquiring, they count as maintaining. An offensive rebound cancels out a missed shot. Evan's model, for one, does the same thing, although it values missed shots and offensive rebounds at a different level than wins produced. I think it kind of has to be that way.

[mystic]

Berri's basis is offensive and defensive efficiency. He is using simple PTS/Poss in order to determine that. Poss is either PE or PA, depends on wether it is offense or defense. PE means possession employed, PA possession aquired. Those two things are basically the two sides of the same coin. If something has a negative sign in PE it just means that it is not counted as possession employed, but quite the opposite. So, point being, offensive rebounding is "possession aquired in Berri's world" on an individuell basis.
Now, if a player scores a field goal after obtaining an offensive rebound, he used no possession according to Berri. I actually just created a scenario in which Berri's formula isn't defined at all. And that is the case for a typical scenario in a basketball game.
On a team level that works fine, because someone else took a shot before an offensive rebound is available, on an individuell level it is not necessary to shoot the ball in order to grab an offensive rebound. That is an important point in order to understand the issue of Berri's work. And that is also what I meant with scoring points without using a possession. By using such formulas Berri is also setting the break even points for two point field goals, 3pt field goals and FT. In fact, a player breaks even in free throws, if he makes 45% of his FTA. Well, using the rounded marginal values of Berri it becomes even 44.1%. And that all is setup via the used formula for team possessions "employed".

And that's how a player like Michael Smith can end up being the 7th best player in the league according to WP48 (only players with significant minutes counted) in 2001 and not getting any contract in the NBA again at the age of 28.

Berri can tweak the numbers like he done it now with defensive rebounding, but the metric will still be pretty useless. He can change the way he calculates PA and PE, he could just use Dean Oliver's formulas for individual players here, but that would obviously a bit more complicated. And yes, that is the only way to adress the real issue of Wins Produced.

[Mike G]

Landry Fields is better than Nowitzki, Kobe, Duncan, etc ... Is that because someone has said he is a guard? And as a guard, he's the best rebounder in the league, by a big margin.
But if he's a forward (a la b-r.com), he's an average rebounder. Does his rating change, as though he's not as good?

And that's how a player like Michael Smith can end up being the 7th best player in the league according to WP48 ...

Are you saying Dennis Rodman wasn't the best player in the world, right up to the moment he was unemployable?

[DSMok1]

Landry Fields is better than Nowitzki, Kobe, Duncan, etc ... Is that because someone has said he is a guard? And as a guard, he's the best rebounder in the league, by a big margin.
But if he's a forward (a la b-r.com), he's an average rebounder. Does his rating change, as though he's not as good?

Yes, his rating would change substantially, due to the position adjustment. He'd still be rated highly, because he doesn't handle the ball much and so has few turnovers, and also shoots infrequently, and thus has a higher percentage.

There's still the issue with all shots being compared to league average, rather than a "replacement level" shooting percentage. If you shoot, say, a hair under league average, the more shots you take the lower you will be rated. (Whereas if you compare to a replacement shooting percentage, which accounts for the marginal value of shots, a player with a ton of shots at just below league average percentage is quite valuable, because then the other players on the team can take fewer, higher-percentage shots and the overall team percentage will go up.)

[mystic]

In order to break even in WP48 on 2pt shots, a players has to convert 51.5% of his attempts. The league average is around 49%.

As I pointed out the real problem comes from the used formulas. Berri is treating individuell players like teams, there is basically no interaction happing according to Berri's WP. When a player takes a field goal attempt the possession for the team is not gone in 100% of the cases, but in Berri's formula a player just used a possession which will never come back again unless he can grab an offensive rebound. On the team level that works fine, a someone from the team shoots the ball, someone else might get the offensive rebound, no possession was lost. But that is not working for individual players. It is no surprise that Oliver's formula for individual total possessions is more complicated than FGA+0.45*FTA+TO-ORB. And that formula is used to determine offensive efficiency for an individual player. Just look at Michael Smith anno 2001, the guy scored 301 points, but only used 177 possession for that. Well, in reality Smith was a 52.7 TS% player and had a 13.8 turnover rate, hardly the offensive juggernaut Berri's WP48 thinks he is. And that screws up the whole metric. It can't be fixed by just tweaking some factors here and there. Tweaking the defensive rebounding numbers is just trying to cover up the effects, not actual getting rid of the cause.

Btw, Jason Richardson was as valuable as Dirk Nowitzki last season according to WP48. Who would have thought that? And the position adjustments leads to the interesting result that Tracy McGrady had even a higher WP48 last season. In fact according to WP overall the Dallas Mavericks should have asked the Detroit Pistons for a trade, Greg Monroe+Tracy McGrady for Nowitzki+Terry, that would have given the Mavericks more wins overall.

[xkonk]

Now, if a player scores a field goal after obtaining an offensive rebound, he used no possession according to Berri. I actually just created a scenario in which Berri's formula isn't defined at all. And that is the case for a typical scenario in a basketball game.

Evan, your system works the same way, right? If player A gets an offensive rebound and later makes a shot (ignoring that you consider assisted and unassisted shots differently), he would have earned points without using a possession, correct? But the team would have used a possession because player B took the shot that player A rebounded. Perhaps you could explain this in a way that has nothing to do with Wins Produced.

[EvanZ]

In ezPM:

Player B loses 0.7 pts for missing a shot. The defense gets +0.7 pts.

Player A gets 0.7 pts for getting an OREB. The defense (his counterpart) loses 0.7 pts.

Player A then gets 1 pt for making a 2-pt shot or 2pts for a 3-pt shot. The defense loses 1 or 2 pts accordingly.

Every offensive result is always balanced by a defensive result.

[Crow]

The defensive rebounding adjustment seems like a good move to me. It may not be the perfect defensive rebounding adjustment for everyone or the only move desirable move (as already noted the shooting efficiency break even point might be too high and offensive rebounds might also still be too high for some) but it lessens or removes a big issue.

Instead of 6 bigs in the top10, it is just 3. That in itself is fairly significant to me.


Berri says: "...86% of the variation in Win Shares per 48 minutes was explained by the player statistics. That tells us that Win Shares has a much larger adjustment for team statistics. Again, not sure I have ever seen anyone troubled by this feature of Win Shares. Again, one suspects that this is because the on-line stats community doesn’t devote much effort critiquing the models generated by members of the on-line stats community."

I've noted the large undifferentiated impact of team shot defense for all minutes (regardless of whether the player is on the court) on individual Win Shares probably at least a dozen times here and other places. Sometimes I have also addressed the issue of the awarding of all the credit to individuals for successful steal and block attempts and the lack of possibly appropriate (based on what happens next) individualized penalty for unsuccessful attempts. Berri is either unaware of or ignoring such comments. Not that this matters much to me at this point but it is still seems worth a brief mention. Lots of other folks are aware of these issues with WinShares, even if they don't highlight it.

Critique of perceived weaknesses has not been equally spread among metrics, but most metrics have had their turn of getting some, usually when they were fairly new. There has been some critique of WinShares and certainly some of APM / RAPM too. PER has also been critiqued for what it lacks and the lack of update. While I think EZPM and the other play by play based metrics Evan has produced go beyond most or all simple linear metrics, I have poked at it and offered some critique along the way too.

Going forward I hope the strengths and weaknesses of metrics are noted fairly and with less rancor.

[mystic]

Evan, your system works the same way, right? If player A gets an offensive rebound and later makes a shot (ignoring that you consider assisted and unassisted shots differently), he would have earned points without using a possession, correct? But the team would have used a possession because player B took the shot that player A rebounded. Perhaps you could explain this in a way that has nothing to do with Wins Produced.

I don't see Evan counting possessions for individual players in order to calculate offensive efficiency. As I see it he just counts marginal values. Berri's system is based on offensive and defensive effiency. In fact, even though Berri is claiming that he used regression to get his results, he just did one regression with offensive and defensive efficiency as independent variables. Everything else is based up on league averages and the formulas for offensive and defensive efficiency.
As I pointed out the issue comes from using team related formulas for this. While you can count the offensive possessions of a team by using FGA+0.45*FTA+TO-ORB, you can't do that with individual players anymore. For sure, you can break the team numbers down to the individual players, then apply some arbitrary parameters to each variable. If you then sum up all players, you will get each time a pretty close correlation to the scoring margin of the team. But that doesn't mean that you actually evaluated players here. In fact, that system is handling players as if each of them would be a team by himself. And that's why the results are looking so fishy. Everyone who has above average offensive rebounding numbers (most times just due to positioning in the offensive set, not so much about skill), but takes below average amount of shots while not handling the ball much anyway, will have a really good WP48 value. The reason is the definition of the offensive efficiency, not the rest of the calculation.

WP48 overrates low usage players who can grab offensive rebounds. It is like the opposite of PER, a metric in which high usage players are getting overrated. And that's why both of those metrics finished in my lineup check in a similar fashion and were substantial worse than WS48 and my own metric.

[Mike G]

"... one suspects that this is because the on-line stats community doesn’t devote much effort critiquing the models generated by members of the on-line stats community."."

Does this mean DB et al do not consider themselves to be " members of the on-line stats community"?
Because this stuff has certainly been critiqued, here and elsewhere.
And is that not the genesis of the problem? -- That in re-inventing the wheel, oblivious to the models already out there, they designed a basically square wheel, enshrined it in a book, and thereafter have had to defend the book?

Economists are pretty notorious for bygod sticking to their theories, damning all evidence (but some), plodding forward with the same failing applications. Since they all have entirely separate descriptions of what has happened and why, they can say anything whatsoever, and no one can possibly know if it's true, the exact opposite of true, or just a political expediency. Many economists are full-time propagandists.

This of course makes the job of a truth-seeking economist much more difficult. He has to un-teach all the propaganda. So the world economic situation is handled by a bunch of experts behaving randomly.