APBRmetrics

[Conan the Librarian]

Thank you, asimpkins. That is exactly what I was looking for (that or feasable explanation from Berri or a WoW supporter).

APBRmetrics: helping fools like me understand simple statistics since...whenever this forum opened.

[Statman]

[Mike G]

Can anyone offer a simplified explanation for why players would rank better or worse depending on what position they are supposed to play? This seems to add yet another wrinkle in the WoW -- just in case it isn't implausible enough?

Does anyone besides myself think we actually haven't been overly cruel to Berri or his adherents who represent him here? Have we actually been too generous and/or patient? Flint, I have to say, represents the most ardently devoted fan of another's metric, that I have seen. He also is least able to defend it.

That's why I think Flint and Berri are one and the same. I've never encountered such an adherent; you couldn't invent him. He's got to be an alter-ego.
_________________
`
39% of all statistics are wrong

[kjb]

[Guy]

Here’s a thought experiment that I think sheds some light on the question of valuing rebounds vs. scoring in Win Score/Wins Produced. It won't break new ground for most here, but I think it may be helpful for explaining the problem to those on the fence (if anyone is still there).

Suppose that the NBA had decided years ago to have scorers credit a “missed rebound” (RBMiss) for each rebound by an opposing team, defined as the player judged to have had the best opportunity to get the ball on the non-rebounding team. So “Rebound opportunities” (RBO) = RB + RBMiss, and for each player we knew their RB% (RB/RBO) which averages 50% (c. 70% on defense, 30% on offense). However, this NBA decided it was not important to record FGAs.

In this parallel universe, the WOW authors would run their regressions and discover this Win Score formula: Points + Rebounds – .5*Rebound Opportunities + Steals + .5*Assists + .5*Blocks – TO - .5* FTAs - .5*PF. (That is, RBOs replaces FGAs.) The authors got a lot of attention for pointing out that rebounding was overvalued in the NBA, and identified several players with impressive raw rebound totals but who actually were hurting their team because they rebounded at less than a 50% rate. Critics objected that some top scorers, like Allen Iverson, were rated too highly because no adjustment was made for scoring opportunities consumed. But the authors argued that in the absence of clear proof that some scorers took more shots than others, we must give credit to those who scored the points. Plus, the model predicted 94% of the variance in team wins!

This if of course simply the WOW story in reverse. Everything WOW says about why we shouldn’t just assume players with high point totals are great players applies exactly to how rebounds are handled in Wins Produced. It assumes that rebound opportunities are distributed randomly (except big men have more than guards), so that every rebound above average represents a net additional rebound to the team. But that’s no more plausible for rebounds than for scoring (where correlation of PTS and FGA is over .8). We don’t know what the correlation is between RB and RBOs (because RBOs aren’t recorded at the player level), but there is strong evidence that the correlation is very high:

* the SD for RB/48min at the player level is around 3.8, but at the team level is just 1.4 -- actually lower than for individual players. If each player’s rebounds were independent of his teammates’ opportunities, then the team SD would be about 8.5, or 6 times as big.

* the correlation between RB/G for a team’s top rebounder and RB/G for his teammates (2006-07) is around -.8. Clearly, many of the RBs by the top rebounders are just taken from other players. It’s not a zero-sum game, of course, but much closer to that than to win-win.

These results tell us there is a huge negative correlation between a players’ rebound totals and that of his teammates. And that means WP should value rebounds much lower -- perhaps .2 or .3 -- rather than 1.0.

[HoopStudies]

[basketballvalue]

[Guy]

[schtevie]

There's a fellow going by the name of T.G. Randini who is currently dogging Prof. Berri. in the comments on his blog.

The there and back (mostly the there) is quite amusing, and I was just wondering if anyone here want to fess up to another alias.

[HoopStudies]

[Guy]

I agree with all that, including your emotion/logic point. Position adjustments can certainly make sense. But if have an average guard (G) and an average center (C), who you believe are equally valuable, but your metric tells you C = 2*G, then you should think about why that is.

One possibility is there is value outside your model. A 1Bman produces about 25% more runs than a catcher in baseball. Why? Defensive ability that isn't measured by offensive metrics. So if that's the answer here, it seems the WOW authors should be interested in trying to understand the 50% of guards' value they currently aren't measuring. (Of course, they can't be missing 50% since they've explained 95% of team wins, but let's leave that aside...)

The other possibility is that the metric isn't measuring value properly. In which case it needs to be modified.

Whichever it is, the discrepancy should be dealt with analytically, not just defined away. At least, that's my (unemotional) view....

[Analyze This]

[Harold Almonte]

I think the position adjust is a not so unfair way to try to make up for usage. But as Dean said positions are not only poorly defined, there are a lot of difference at both ends, among a lot of players defined as at the same position. I think the best way would be to split players according to their usage, and position is the more similiar thing that easiness can give us (you know C,PF= more R,BLK; PG= more AST,TO; Shooters= more FGMissed and FGA, etc). Something mechanical.
If you could define players by usage, you can have different break evens for each usage level, and is more fair than to say a 50%-2 shots player is better than a 40%-25 shots, because 50% is supposedly the same dificulty level for both and everybody, shooting from everywhere and every way. And the same can be applied to RebR%, ballhandling (A/TO), and whatever other action rate.
But, i think everybody agree that if one action is rated against misses or attempts, all actions inside a rating must be rated the same, and every action need an efficiency, and a not zero break even, (preferably not by position, but by stat usage).

[Flint]

Posting anonymously on a website seems out of character for an economics professor. I am definitely not Dave Berri.

While my burning desire ot discuss the WOW here has been thoroughly quenched, I do think it would be very interesting to see how consistent player performance is, year to year, across all available metrics.

[DLew]

Rebounding is very consistent year to year. Throw a team adjustment on there and you will get something that explains wins as well as Wins Produced and is more consistent.

Also, no offense Dean, but I think the discussion of WOW in this thread recently has been quite civil and logical. Guy has been making some really excellent critiques that shouldn't be dismissed. There's not much new to say about WOW that hasn't been said, but I thought that asimpkins' summary of the situation was quite good:

"Defensive rebounds are credited as gaining an entire possession and shooting is penalized as losing an entire possession. This is way off, and it perfectly explains the bizarre results... Why would someone even think these weights were correct in the first place? Regression. Econometrics. Right. But... whatever Berri may have correctly calculated at the team level can't necessarily be applied at the individual level. The mistake has an understandable miscalculation behind it... The results don't make sense. The problems are an obvious result of the weights assigned. There's a good explanation why the weights are incorrect."