Why don't advanced stats like Bradley's defense?

[kozlodoev]

He has a negative Defensive Boxscore Plus Minus.  His DRPM is negative also, and ranked quite low leaguewide.  He has 2.6 Defensive Win Shares on the season, which is fine, but not close to top 20.  His defensive rating is 106, which is not horrible, but not particularly great either.  His steal percentage of 2.2 is fine but far from exceptional.

Part of it is that these metrics are awful measures of defensive ability. Another part is that his defense is somewhat overrated.

You are right about #1 but the debate comes from both Lillard and McCollum stating that they believe he is the best.  And they would know much better than anyone on this blog.

The fact that Lillard and McCollum doesn't like having someone in their face when they dribble the ball doesn't necessarily make them an overall impact defender.

What about guys like Steph Curry who praised Bradley's defense?

You don't really understand the core of the argument, do you?

I think I should elaborate, personally while Bradley is an excellent on ball defender I believe that defensive stats rate him poorly because plays alot with the bench. It's very tricky measure defensive worth which is why I often go with the eyetest and from what I see Bradley guards the elite point guards as well as any point guard. So I can see why the opponents praise Bradley's defense over the other guards

My impression is that folks tend to go gaga over Bradley because of his pressure defense in the back court. While this is flashy and potentially annoying to opposing guards, it's certainly not the main factor in determining whether one is an impact defender. Stuff like being able to stay in front of the ball in the half court, where you angle the guard in set plays, how you handle the pick and roll, how you close down on jump shots, etc. are perhaps less evident, but decidedly more important.

Bradley is certainly solid, but branding him the top defender in the NBA because he was once getting away with antics that are remedied with a solid  back court pick is a bit over the top.

[Snakehead]

Defensive stats are largely bunk. They don't adequately take into account things like rim protection. If Avery gets best inside, he gets penalized by a likely made hoop. If Dwyane Wade gets beat, Hassan Whiteside cleans up his mess.

Similarly, a lot of RPM is a fancy formula for +/-. This rewards players who are clearly better than their backup. However, we are a deep team with good defenders. If Smart subs in for Avery, the defense doesn't lose a step, so the RPM formula says his defensive impact isn't significant. That's not true, of course.

This is how I feel about.  I'm all for the stats and that matters but the eye comes into it, especially on defense.  I think Bradley, Smart, and Crowder are all top level defenders.

[Snakehead]

In other words, let common-sense be your guide for defense.  Part of common-sense is listening when the games best players say that someone is good at defense.  They all know who they can score on and who they can't.  If most everyone is saying the same couple of players, those are the good defenders.

http://www.basketball-reference.com/about/bpm.html

I also put a lot of credit into what the players and coaches say too.  It tells you a lot.  For example heard a player talking recently about how he thinks DeAndre Jordan is such a great defender because he actually is a KG type communicator/QB of a defense and is constantly talking.  Not something I think many fans talk about with him but the players know and give him credit for.

EDIT: I remembered it was actually Micah Shrewsberry talking on the Globe's Celtics podcast, which was a good listen.  He put him up there with Paul, LeBron, and Rondo as far as communication and reading an offense.

https://soundcloud.com/user-345535714/no-18-celtics-assistant-micah-shrewsberry-1-on-1

[chilidawg]

Isn't smart pretty weak in advanced defensive stats as well?

He's 6th among PGs in DRPM. I'm still so shocked that Avery is 47th...

The error bars on DRPM are so gigantic, that's basically saying that there is no difference.

Can you explain this more?

Rather than me just repeating explanations that have been well stated by others, I recommend the following Wages of Win article, which goes into some of the issues with adjusted-plus-minus methods at length:

http://wagesofwins.com/2011/03/05/deconstructing-the-adjusted-plus-minus-model/

The net takeaway is that you should take RAPM with a gigantic grain of salt.

After reading about the particular RPM that ESPN uses, with priors that include height and age, I'm good and ready to forget about it in favor of a better plus-minus metric, maybe one that's blended with a box score metric like some people have been clamoring for. I generally trust plus-minus metrics, and ESPN's just happens to be the most convenient to look up. But man, when you're docking rookies just for being rookies, with zero regard for individuality...when your metric gives players a boost just for being tall...nah, I need something better. The height thing, which I had always thought was literally a joke whenever I saw it mentioned, is probably a significant reason why Bradley's ESPN DRPM is so low.

Where did you see the part about height and age? 

I read the article and readily admit about half of it was over my head.  But two things stuck out.  One was that he clearly says he's looking at one particular APM model.  Second is that the article is from 2011, which is a couple years before ESPN came out with their model.  Maybe they've got a new approach, maybe not, I don't have any idea.

To M's point about collinearity, if Crowder and IT are on the court together so much, why are their DRPM's so different?  Because IT is short?  It can't be that simple can it?

This is stuff I'm always curious about, so I'm not arguing with you, just trying to understand.

[PhoSita]

In other words, let common-sense be your guide for defense.  Part of common-sense is listening when the games best players say that someone is good at defense.  They all know who they can score on and who they can't.  If most everyone is saying the same couple of players, those are the good defenders.

http://www.basketball-reference.com/about/bpm.html

I also put a lot of credit into what the players and coaches say too.  It tells you a lot.  For example heard a player talking recently about how he thinks DeAndre Jordan is such a great defender because he actually is a KG type communicator/QB of a defense and is constantly talking.  Not something I think many fans talk about with him but the players know and give him credit for.

I haven't watched Deandre very much (I detest the Clippers, honestly), but I have a feeling he's improved A LOT in these hard-to-notice areas that you're pointing out.  I suspect Doc's tutelage has paid dividends here.

[Boris Badenov]

To M's point about collinearity, if Crowder and IT are on the court together so much, why are their DRPM's so different?  Because IT is short?  It can't be that simple can it?

Collinearity doesn't mean that the estimates must be similar.

It means that even if the estimates are different, that difference may not be "statistically significant,"  meaning you can't within a reasonable margin of error rule out that something else is true (in other words, IT's and Jae's DRPMs could reasonably have a wide range of other values).

These values we see for APM and BPM are statistical estimates, and should be treated as such, but they rarely are, probably because doing so would reveal how messy/uninformative they can be.

In contrast, just as an aside, a player's net plus-minus is not a statistical estimate - it's something one can know with certainty. While net plus-minus a bit more flawed theoretically, it's less susceptible to noise in actually measuring it, as these other measures are.

[kozlodoev]

To M's point about collinearity, if Crowder and IT are on the court together so much, why are their DRPM's so different?  Because IT is short?  It can't be that simple can it?

Collinearity doesn't mean that the estimates must be similar.

It means that even if the estimates are different, that difference may not be "statistically significant,"  meaning you can't within a reasonable margin of error rule out that something else is true (in other words, IT's and Jae's DRPMs could reasonably have a wide range of other values).

These values we see for APM and BPM are statistical estimates, and should be treated as such, but they rarely are, probably because doing so would reveal how messy/uninformative they can be.

In contrast, just as an aside, a player's net plus-minus is not a statistical estimate - it's something one can know with certainty. While net plus-minus a bit more flawed theoretically, it's less susceptible to noise in actually measuring it, as these other measures are.

Collinearity is a fancy way of saying that if you use corellated variables as explanatory variables in a regression, the statistical power of your model goes down the toilet. Collinearity doesn't imply that the variables are numerically similar, just that one can be used to predict the other with a great degree of accuracy.

This is all just tangentially relevant here, because this "model" is data mining of the most brainless kind. It seems that with the advent of computing power and "big data", people have forgotten that quantitative analysis serves to test a hypothesis about how the world works, not poke around for spurious correlation.

[Dino Pitino]

Where did you see the part about height and age? 

My bad, I did a little googling after reading that article. The math dudes who reverse engineered ESPN's metric say that height and age are priors in Engelmann's Bayesian formula. (Or whatever, I'm not a math dude, lol.) Isaiah's RPM is influenced by the fact that shorter players are assumed from the get go to be worse. Rozier's is influenced by the assumption that first year players are worse. Etc.

[kozlodoev]

Where did you see the part about height and age? 

My bad, I did a little googling after reading that article. The math dudes who reverse engineered ESPN's metric say that height and age are priors in Engelmann's Bayesian formula. (Or whatever, I'm not a math dude, lol.) Isaiah's RPM is influenced by the fact that shorter players are assumed from the get go to be worse. Rozier's is influenced by the assumption that first year players are worse. Etc.

A "prior" probability is not an assumption, it is a known prevalence rate of something.

[PhoSita]

Where did you see the part about height and age? 

My bad, I did a little googling after reading that article. The math dudes who reverse engineered ESPN's metric say that height and age are priors in Engelmann's Bayesian formula. (Or whatever, I'm not a math dude, lol.) Isaiah's RPM is influenced by the fact that shorter players are assumed from the get go to be worse. Rozier's is influenced by the assumption that first year players are worse. Etc.

A "prior" probability is not an assumption, it is a known prevalence rate of something.

In this case, isn't it an assumption that a general trend applies to a specific player?

[kozlodoev]

Where did you see the part about height and age? 

My bad, I did a little googling after reading that article. The math dudes who reverse engineered ESPN's metric say that height and age are priors in Engelmann's Bayesian formula. (Or whatever, I'm not a math dude, lol.) Isaiah's RPM is influenced by the fact that shorter players are assumed from the get go to be worse. Rozier's is influenced by the assumption that first year players are worse. Etc.

A "prior" probability is not an assumption, it is a known prevalence rate of something.

In this case, isn't it an assumption that a general trend applies to a specific player?

No, it's an estimate based on the best available data - in the absence of any further knowledge about the player.

[Alleyoopster]

A good part of it may be due the reputation he had when he came into the league. As a rookie and his second year he displayed uber athleticism and was incredibly tenacious on defense. Working his body to extremes lead to injuries. Finally, he woke up and realized that if he kept playing at that intensity he wouldn't be in the league very long.

He still plays good defense, but not to his rookie year level. He knows that if he pushes his body he'll be back on the sidelines.
Thus, part of this discussion may be from a distorted vision many of us still have of him based on how he played early in his career.   

[PhoSita]

A good part of it may be due the reputation he had when he came into the league. As a rookie and his second year he displayed uber athleticism and was incredibly tenacious on defense. Working his body to extremes lead to injuries. Finally, he woke up and realized that if he kept playing at that intensity he wouldn't be in the league very long.

He still plays good defense, but not to his rookie year level. He knows that if he pushes his body he'll be back on the sidelines.
Thus, part of this discussion may be from a distorted vision many of us still have of him based on how he played early in his career.

I thought much the same way, but the comments this year from top players calling him one of the best perimeter defenders in the league gave me pause.

I think there really is something to the idea that opposing ballhandlers might rate (perhaps overrate) his overall defensive impact higher than coaches / players elsewhere on the floor might, because subjectively that's got to be really annoying. 

But the fact that Bradley makes things feel tougher for some of the top ballhandler scorers in the league says something about his defensive ability.

[Vox_Populi]

Collinearity is a fancy way of saying that if you use corellated variables as explanatory variables in a regression, the statistical power of your model goes down the toilet. Collinearity doesn't imply that the variables are numerically similar, just that one can be used to predict the other with a great degree of accuracy.

This is all just tangentially relevant here, because this "model" is data mining of the most brainless kind. It seems that with the advent of computing power and "big data", people have forgotten that quantitative analysis serves to test a hypothesis about how the world works, not poke around for spurious correlation.

Unless I'm misunderstanding, this sounds like you don't think much of RPM in general. Am I wrong?

[PhoSita]

spurious correlation.

This should be the tag-line of ESPN as a whole.