Nate Silver is one of my personal heroes. He’s a statistical rockstar. The work he’s done in baseball and in politics is truly amazing and an inspiration. I’ve been reading his stuff for years.
I really can’t believe I’m about to take the piss at Nate Silver. But politely disagreeing with a fellow stats person is fine I guess (and I won’t even bring up anyone’s mother). But I disagree with Mr. Silver, and think he got it wrong.
But I’m skipping the lead. How did I get here? First off there’s this piece:
The Nets and Knicks May Be Better Off Without ‘Melo.
In it Jared Diamond of the Wall street Journal (with an able assist from our own Prof. Berri) debunks some of the hype surrounding an old favorite, Carmelo Anthony (did you hear he might be traded?).
The relevant bulletpoints from the Professor:
- The latest rumored New Jersey deal (Derrick Favors and Devin Harris to Denver, The Nets get Chauncey Billups and Rip on top of Melo) would only make the Nets a 30 win team.
- If the Knicks give the Nuggets what they want (Fields, Chandler and picks) for Melo they would win roughly 29 games over a full year (and I would laugh myself sick ).
And I agree with both of these assessments. As you might imagine, a lot of people do not.
One such person is the aforementioned Mr. Silver in this lovely piece:
Why Carmelo Anthony Is the Ultimate Team Player (and What ‘Advanced’ Stats Miss About Him)
His claim is simple. Advanced stats such as Wins Produced (see here for the Basics) miss the true value of Carmelo Anthony as a player. To quote:
In taking all of those shots, however, Anthony has also done something else: he’s made his teammates much more efficient offensive players.
And because he makes his teammates more efficient by being inefficient himself, Melo is more valuable than he appears. Now Mr. Silver does prove his first point but as we shall see his second point fails when tested. Let’s talk data.
I went through the trouble of pulling every single game for the nuggets for the past two years (see here ). I then divided all the games into games where Carmelo played and games where he did not. Finally I worked out Points generated per shot for each scenario. The results are here:
You can see that Mr. Silver is right. Melo’s teammates are more efficient with him on the floor . But only slightly (.9% this year and .1% last year). The problem though is that as a Team they are more efficient with him off it. Because Melo is inefficient and he takes a lot of shots which more than eat up the marginal gains in offensive efficiency he get’s his teammates. In Melo’s defense his team is better off with him on the court (but that has to do with other stats and who his replacement is).
Melo is a good but not even close to great player but the appearance and not the substance game makes him seem like more than he is. So much so that even those who should know better are beguiled and bewitched to defend him. Anyone who’s thinking of buying what Denver is selling needs to remember the numbers and beware of his eyes leading him astray.
some dude
01/16/2011
I agree that Silver’s analysis was short-sighted. 82games.com has the on-off stats. 2 years ago, when they made it to the WCF, the team shot better when he was on the bench (efg%). It was better offensively, however, because it shot more FTs and committed less turnovers with him on it. And of course his backup wasn’t nearly as good
I’ve always maintained Melo is overrated by some and underrated by others. I also don’t think the Knicks should trade for him.
That said, a Felton-Melo-Amar’e team is not winning less than 35 games (if they all remain healthy). Just another case of WP not being able to properly predict the future.
Not sure about the Nets because that seems like an odd mix. For the life of me, I can’t figure out why Melo would want to play with Billups and Rip 5 years too late. I believe the agents are pushing the Rip to NJ thing, not Melo.
I actually agree with Sports Guy that he should go to the Clippers (though his argument is hypocritical…). Well, other than not wanting to play for Sterling, of course. Pretending the team was stolen from him, that would be where I would try to go if I was him.
EvanZ
01/16/2011
What about Melo on Chicago? He’s never been known for his defense, but neither was Rose until Coach T arrived. That team is not especially efficient on offense either. My guess is that he would have a net positive effect.
some dude
01/16/2011
It depends. If we’re talking about Deng + Taj Gibson, I agree. If they’re giving up Noah, I don’t like it.
We haven’t even seen Boozer-Noah with both in rhythm yet this season, either. I think Noah’s ability to defend his man and be really good in help D coupled with his ability to pass makes him a unique asset that can really help win you a title.
EvanZ
01/16/2011
No, I definitely meant Deng for Melo.
Shareef
01/16/2011
“Just another case of WP not being able to properly predict the future” I could not agree more but what is crazy is how this stat’s devotees remain so cocksure with none of the humbleness that you find among baseball metricians. Soon we’ll see an “explanation” of “increased productivity” fueling the post trade Orlando Magic. I’ve seen every Knick game, Landry Fields is wide open. He would not be if David Lee was on the floor instead of Amar’e, and many of his rebounds would have landed in Lee’s hands, and he would be ordinary rather than “extraordinary”. Carmelo has the ability to fill the void that needs filling on any team. He is asked to draw double team and get the opponents front line in foul trouble he does a great job of it.
bduran
01/17/2011
Adjusted +- also says that Fields is the NYs best player, but I mean, you’ve seen every game.
Mike
01/17/2011
How do you know that?
I mean, compare the two teams: http://nerdnumbers.com/automated-wins-produced 2011 and 2010, and the big differences are Felton and Fields, much more so than Amar’e for Lee.
Shareef
01/18/2011
You guy’s like Landry and so do I! The difference between you guys and me, Landry Fields, and Landry Fields’s mamma is I don’t think he’s better than Kobe Bryant, but um it’s all here in the chart http://nerdnumbers.com/automated-wins-produced As my granny used to say, “What you say listens fine, but still I’ll be havin’ my own figurin’ if you don’t mind.”
arturogalletti
01/16/2011
SD,
Not saying Melo isn’t an above average player. I agree with you on the 35 wins (that’s why I’d be laughing, because 35-45 wins and no cap room is not an upgrade over 45-55 with cap flexibility).I’d also worry about dim.returns with melo & amare. For it to work, Melo would have to change his game.
The Nets would be a car wreck. Rip would sulk.Chancey would be looking for a release.
The Clips would be a great fit actually.He’d actually make that team better.
EvanZ
01/16/2011
If the Knicks traded Fields and Chandler for Melo and JR Smith, the WP would almost break even. It would be interesting to see.
arturogalletti
01/16/2011
Evan,
I agree.I just don’t think it happens.
Neal Frazier
01/16/2011
Is there a reason to prefer PPS over TS% when doing an analysis like this? Is one of them more closely corrolated (insert R v R^2 arguement here…) to WP? I ask because the nearly 4% increase in TS% that Mr Silver notes seems like a big jump and is backed up by a big sample size and its just odd that it disappears almost completely when viewed by PPS, which I always assumed mirrored TS% closely.
Cheech Cohen
01/16/2011
Neal has a good point. I think many will agree that TS% is the better statistic to use here.
arturogalletti
01/16/2011
It’s just personal preference.I like looking at Points generated per fga (and treat free throws as free :-)).The data set is built so I can re-run it easily enough, gimme a couple of hours and I’ll post it.
ilikeflowers
01/16/2011
Arturo, what happens in general (not accounting for quality of backups which is critical as well) when you remove a high minutes, high scoring, low assists player from a team’s lineup?
I have no problem assuming that a high scoring player will draw a defense’s attention away from other players acting as a kind of team wide assist (even though some coaches will focus on defending the teammates instead of a high volume inefficient scorer). But the size (relevance) of the effect is important and looking at one instance isn’t going to tell us much about the general impact and thus we don’t know how often to expect The Carmelo Effect in the NBA just due to randomness.
arturogalletti
01/16/2011
ilf,
I’m also curious about this. I’d need to go out and build the data sets. I suspect that you will see and effect but that it’ll only matter if the volume shooter’s efficiency is greater than a minimum number (i.e so he doesn’t shoot the gain away). So going by that chart I put up on friday, Lebron probably has this going for him (Chris Kaman & CJ Miles, not so much).
Guy
01/16/2011
Nice post, Arturo. Melo’s impact on his teammates’ efficiency does appear to be quite modest.
The improvement in team WP with Melo though is fairly significant — about .020. With Melo using about 16% of MP, that implies a difference between him and his replacements of about .128 WP48. As it happens, that’s also his WP48 over these seasons. That would all add up if his replacements have been .000 WP players overall. But if his replacements have been, say, .050 players, then it would appear he increases team WP more than expected.
BTW, have you ever looked at teams’ 5-man units, to see how well their individual SP48 match up with the unit’s WP48? That could be pretty interesting. Then to the extent certain lineups overperform or underperform (if any do), you could look to see what specific characteristics those units seem to have.
arturogalletti
01/16/2011
Guy,
I really did not do a lot of controls on the data (schedule,opponent,rest) since I was going at a particular number.
It’s an interesting question given what we now know about WP and point margin. I’d be reluctant to tackle as of yet though. We can now look at game to game automatically but minute to minute is still a ways away (translation: the data build would be very,very labor intensive). That said, I suspect defense and opponent performance would drive a lot of the variation.
arturogalletti
01/18/2011
Guy,
I have an idea of how to test this at least initially. I’ll get around to it in the next few weeks.
Guy
01/18/2011
I was thinking of something pretty simple, just comparing the actual point differential for each unit to the predicting differential based on the sum of the players’ WP48. For example, top three 2009-10LAL units in MP, with point differential (assuming 95 possessions), were:
1 Fisher-Bryant-Artest-Gasol-Bynum + 11.4
2 Fisher-Bryant-Artest-Odom-Gasol +3.8
3 Fisher-Bryant-Artest-Odom-Bynum +4.8
Using seasonal WP48, we get these predicted differentials (WP48/differential):
1 .743 / +7.4
2 .834 / +10.1
3 .697 / +6.0
Now we don’t expect perfect predictions even from a perfect metric, due to small sample size. But if you look at units that overperform WP, as a group (large combined sample size), then you could see if they have any distinct characteristics (good or bad rebounders, lots of assists, whatever). Do the same for underperforming units. And if the only difference between actual and predicted performance is noise due to sample size, there shouldn’t be any patterns in the data.
arturogalletti
01/18/2011
Guy,
When doing lineups, you have to account for the fact that they’re not playing baseline opposition. Starters would actually expect to see >.100 WP48 on the other side. It get’s downright complicated fast. As I said I have some ideas around how to work around that using large samples.
Guy
01/18/2011
Sure, lineups face varying quality of opposition. But it’s not like WP is opponent-adjusted in the first place — each players’ stats are treated the same, so in fact all starters are already a bit underrated and bench players overrated. And it would mainly be a problem for the small-MP units, who face well-below average opposition, not when looking at frequently-used lineups. Plus, what you really want to learn is if there are characteristics that lead to units underperforming or overperforming their expectation. I’m sure the lineups that overperform will tend to have faced weaker opponents. But that still wouldn’t cause them to consist of high-assist or low-assist players, or be stronger on possessions than efficiency (or vice-versa), or any other particular quality. So if you find systematic differences between over- and under-performing units, that would still be very important and interesting.
arturogalletti
01/18/2011
It does adjust for opponents (just at a season not a game to game level). I agree that it’s an interesting exercise.
Man of Steele
01/16/2011
EvanZ, I think the WP gap is a bit larger, due to the fact that Wilson Chandler plays more as a “big” or composite forward on the Knicks than he should. In Denver, with Nene, David Andersen, and Kenyon Martin, he would be a pure SF, and the WP numbers would look better for him. Thus, I think Chandler as a SF plus Fields would be significantly better than Melo and J.R. Smith … just my opinion, though.
Man of Steele
01/16/2011
Also, I definitely want to be on record in favor of Melo for Deng + picks (from Denver’s vantage point, that is). In my opinion that would be the basic type of move for any stathead GM: swap an overrated scorer for a comparable player who scores a bit less plus a couple of useful extra players and/or draft picks. It opens up cap space, doesn’t damage productivity, and increases one’s chances in the crapshoot that is the draft. It’s a win-win-win.
arturogalletti
01/16/2011
MOS,
Yeah. The Chicago scenario is just ok for Denver. The Fields/Chandler/Curry/Picks for Melo/Harrington would be fantastic. They get cheaper (after they cut Curry), younger and they get picks.
jamerchant
01/18/2011
Any reason you didn’t use the simpler statistic of points per possession? It seems to me that captures the effect we’re looking for (offensive efficiency) better than points per shot (although I’m sure the two are highly correlated).
And I was a bit confused by this statement: “The problem though is that as a Team they are more efficient with him off it.” How so? By your measure, the team was (slightly) better on O when Melo played, right? And their WP was higher? I’m not trying to be contentious, I just didn’t get that sentence.
Not to shill myself too shamelessly, but I did a similar study of Denver’s game with/without Melo from 2003-11 that found that Denver’s points per possession (offensive efficiency) was 1.10 when Anthony played and 1.07 when he sat: http://jamerchant.wordpress.com/2011/01/18/with-or-without-melo/
That seems to be in line with your points per shot findings, but I took it as evidence that the Nuggets truly were better on offense with Carmelo playing. Whether the team was better is an open question (defense matters, too, of course).
arturogalletti
01/18/2011
Jamerchant,
If you’ve seen my previous work(https://arturogalletti.wordpress.com/2011/01/07/offensive-all-stars/ for example), you’ll know I favor point per possession metrics . In this instance I was trying to keep to the constraint of Nate’s argument (i.e. points per shot or TS%). So even if I use a slightly juiced metric (i.e. biased towards shooting), I still see the proposition as false.
My data shows very similar results.
Guy
01/18/2011
I thought Silver’s analysis was thin, and Arturo presents some good data above. But Pelton’s analysis at Basketball Prospectus seems much stronger that either, given his ability to look at players with Melo on and off the floor in the same seasons. That meets the most serious objections to Silver’s study, which was the failure to control for players’ age, coach, and team. And Pelton finds a VERY substantial Melo effect (though not as strong as Silver reported). In addition, I’d say Dr. Berri’s post lends further support to the case for a Melo effect (though it’s not as strong evidence as Pelton presents). So I think the Melo-improves-his-teammates’-efficiency argument now clearly has the upper hand, unless someone has further evidence to offer.
arturogalletti
01/18/2011
Guy,
I read that piece. It’s a good analysis. But I think the Last paragraph is telling:
“As Silver noted, there is still an important aspect of “fit” that cannot be removed from the numbers. Denver has been able to build a lineup around Anthony’s high usage. The Nuggets have put a series of low-usage, defensive-minded shooting guards alongside him on the wings, for example. With another team already blessed with shot creators, like the Knicks, the value of Anthony’s usage might be limited. In a vacuum, however, a variety of different numbers point to similar conclusions. We may not be able to put an exact number on how many wins Anthony’s ability to create means for his team, but it appears to be larger than zero and substantial enough to make him an All-Star-caliber player despite his other shortcomings.”
There’s still a big issue with multicolinearity (as Melo shares the court with a lot of the same player so is it a Melo or a Chauncey effect which is exacerbated by the fact that this year with Billups out their offense is worse).
And even if I give you the effect, it’s small enough that when Carmelo’s inefficiency is added back in (and his defense), it goes away.
Guy
01/19/2011
I don’t think you can call 7.5 wins a “small” effect. That’s equivalent to raising his WP48 by .137. A better word word be “huge.” And Melo’s defensive shortcomings, while certainly relevant to any team trading for him, is separate from the issue under discussion here — as indicated by the fact that your post (like Silver’s) says not one word about his defense.
And the larger question is whether high-usage players, and perhaps others, have a significant impact on teammates that should be considered in metrics. Here we have a good example of where the answer appears to be yes.