PER and Usage Rate

[PapaG]

Actually, shouldn't those who criticize using these two parameters in their own equation, solely on the basis of 'correlation', actually offer up the data? Minstrel and andalusian both missed the mark badly when they were speculating on the connection.

I wasn't the one claiming "correlation", remember?

 

[blazerboy30]

Actually, shouldn't those who criticize using these two parameters in their own equation, solely on the basis of 'correlation', actually offer up the data? Minstrel and andalusian both missed the mark badly when they were speculating on the connection.

I wasn't the one claiming "correlation", remember?

there appears to be a consistent correlation between a high usage rate and a high PER.

Wait...what?

 

[blazerboy30]

I'm not claiming that I have an extensive data set that I have studied. I am pointing out an observation that I made regarding Usg/PEr, and if meaning can be derived from it. Also, why would I offer an extensive data set for review here?

Well, you started the thread to discuss a "consistent correlation". I would think you might want to offer data to back it up since you claim to be a stat geek.

The primary critic in this thread didn't even know what Usage Rate measured, yet I'm supposed to offer up a data set that would literally take days to plot for a critical review. LOL

I don't know why it would take "literally days" to create a scatter plot with PER on the y-axis, usg on the x-axis and various regression fits to that data. That is a pretty simple task for somebody that is interested.

Speaking of being bitchy...

Don't get all upset. It was an honest question. If you don't know how to write a script to automatically scrape the data from online sources, I was going to offer to help. Sheesh!

 

[MickZagger]

Nerd session!

 

[huevonkiller]

Guys, anyone that wants a good usage-efficiency barometer should also look at this:

http://www.basketball-reference.com/blog/?p=5500

This means that in a 107.5-ORtg environment, the efficiency trade-off for increasing or decreasing usage by 1% is as follows:

Player Type Tradeoff
High Usage (>=23%) 0.833
Mid Usage (18-23%) 1.250
Low Usage (<=18%) 1.667

Basically, after adjusting for the 1995 environment this is the tradeoff for offensive rating and usage. For each extra possession a star takes on, they gain .833 points in offensive rating.

However it does ignore defense, defensive rebounds, steals, blocks, and such. But if you can combine it with an 82games.com approach, it can give you a damn good idea of how a player is performing. For players before 2002 you could look at defensive rating.

I would say that PER is still useful for comparing superstars, taking on extra possessions also has value. Of course so does defense which just as a reminder, don't forget about.

A more recent article about this season:

I realize this isn't always the case for all players -- but as a very general rule it holds, so let's pretend for a moment that this simple model does in fact explain the fundamental usage-efficiency tradeoff in basketball. Under those rules, a player using 18% (or fewer) of team possessions while on the court would see his efficiency change by 1.65 points of offensive rating for every 1% change in usage, a player using 18-23% would see a change of 1.24 pts of ORtg for every 1% of usage change, and a player using 23% or more would see a change of 0.82 pts per 1% change in usage.

http://www.basketball-reference.com/blog/?p=8522

 

[PapaG]

Guys, anyone that wants a good usage-efficiency barometer should also look at this:

http://www.basketball-reference.com/blog/?p=5500



Basically, after adjusting for the 1995 environment this is the tradeoff for offensive rating and usage. For each extra possession a star takes on, they gain .833 points in offensive rating.

However it does ignore defense, defensive rebounds, steals, blocks, and such. But if you can combine it with an 82games.com approach, it can give you a damn good idea of how a player is performing. For players before 2002 you could look at defensive rating.

I would say that PER is still useful for comparing superstars, taking on extra possessions also has value. Of course so does defense which just as a reminder, don't forget about.

A more recent article:



http://www.basketball-reference.com/blog/?p=8522

Interesting info. Thanks for posting! Repped.

 

[huevonkiller]

Also just to clarify, in 2010-2011, for each extra possession a role player takes on they get penalized by 1.65 points in offensive rating. According to skill curves.

Hopefully that clears things up.

 

[blazerboy30]

I pulled the data for all players, and eliminated anybody with less than 300 minutes played.

The linear regression gives: per = 0.54 * usg + 3.8
The r-squared value is: 0.419

Considering that per and usg share many of the same basic stats, it shouldn't be a big surprise that there is a correlation. But the low r-squared value says that much of the scatter in the per data isn't explained by usg%.

 

[PapaG]

I pulled the data for all players, and eliminated anybody with less than 300 minutes played.

The linear regression gives: per = 0.54 * usg + 3.8
The r-squared value is: 0.419

Considering that per and usg share many of the same basic stats, it shouldn't be a big surprise that there is a correlation. But the low r-squared value says that much of the scatter in the per data isn't explained by usg%.

Really? Care to show why you reached that conclusion and actually show the work?

USG and PER share many variables, but they measure different things. LOL

Batting Average and Slugging Percentage use the same basic stats, yet they also measure very different things. Batting Average and OBP use many of the same statistics.

Are you also suggesting that Batting Average and Slugging Percentage have a linear correlation, or that Batting Average and OBP have a linear correlation?

Put your work out so it can be analyzed. Also, a correlation isn't a bad thing if it can be used to explain the variance between to data sets. I just don't see how your conclusion fits anything being posted in this thread. In your case, any player could be plugged into a forumla, and a predictable result should occur. That's not the case. Why not share the p-value of your work, and also share all of the data?

 

[BlazerCaravan]

OBP and SLG are highly correlated... http://cyrilmorong.com/OPS.htm

One problem with OPS, however, is that it adds together two numbers that are highly correlated. In fact, the correlation between OBP and SLG here (only on the hitting side) is .806.

EDIT: I see you've edited your post to replace OBP with Batting Average in most places.

 

[PapaG]

OBP and SLG are highly correlated... http://cyrilmorong.com/OPS.htm



EDIT: I see you've edited your post to replace OBP with Batting Average in most places.

I never posted OBP and Slugging % as being correlated. I added things to my initial post (p-value for statistical significance of the correlation), but you missed on what I posted in terms of the statistics.

Too bad.

The funny thing about them, however, since you brought it up, is that OPS is used as a primary tool for valuing a player under Sabermetrics. Two very correlated statistics being combined to assign value to a player seems redundant, yet it is accepted by a large portion of the Stat Geek community.

EDIT - I see you edited your post to accuse me of posting something that I never posted, and that you also edited away your snarky "Too bad". No need to lie about what I posted, BC

 

[PapaG]

OBP and Slugging being correlated seems to be an argument for using them in a combined sense. Wouldn't this mean that any correlation between PER and usage may actually be beneficial, at least in terms of looking at a similar data set in a different manner, and then combining them to perhaps better utilize both results as an overall value?

 

[blazerboy30]

Really? Care to show why you reached that conclusion and actually show the work?

There is no "conclusion" to reach. It is a simple linear regression. I'm sure you can find the summation equations online or look up the matrix method of least-squares.

USG and PER share many variables, but they measure different things. LOL

Total rebounds and rebound % "measure different things" but it wouldn't be surprising to see a correlation considering they are based on similar basic stats.

Batting Average and Slugging Percentage use the same basic stats, yet they also measure very different things. Batting Average and OBP use many of the same statistics.

Are you also suggesting that Batting Average and Slugging Percentage have a linear correlation, or that Batting Average and OBP have a linear correlation?

I'm not suggesting anything about those stats. But if they are built on the same basic stats, I wouldn't be surprised to see a correlation. You're the one assuming a linear correlation, I have no idea what form of regression would fit the best.

Put your work out so it can be analyzed.

I asked you if you could do it and you got pissy. If you want the data to back up your claims, go ahead and pull the raw data like I did.

Also, a correlation isn't a bad thing if it can be used to explain the variance between to data sets.

Nobody ever said correlation is a bad thing. But you trying to use it to explain the variance in per isn't working, as suggested by the r-squared value.

I just don't see how your conclusion fits anything being posted in this thread.

Damn, dude. You just love trying to be bitchy. You started this thread claiming there "is a STRONG correlation" between per and usg. I posted what the stats say the strength of that correlation is.

You started with a hypothesis about a correlation, and when somebody posts the actual data, you claim you don't see how it relates to the thread?

In your case, any player could be plugged into a forumla, and a predictable result should occur.

Um, no. My data / analysis says exactly the opposite. You might want to look into a stats course.

That's not the case. Why not share the p-value of your work, and also share all of the data?

The p-value is ~0, as expected.

 

[PapaG]

There is no "conclusion" to reach. It is a simple linear regression. I'm sure you can find the summation equations online or look up the matrix method of least-squares.



Total rebounds and rebound % "measure different things" but it wouldn't be surprising to see a correlation considering they are based on similar basic stats.




I'm not suggesting anything about those stats. But if they are built on the same basic stats, I wouldn't be surprised to see a correlation. You're the one assuming a linear correlation, I have no idea what form of regression would fit the best.



I asked you if you could do it and you got pissy. If you want the data to back up your claims, go ahead and pull the raw data like I did.



Nobody ever said correlation is a bad thing. But you trying to use it to explain the variance in per isn't working, as suggested by the r-squared value.



Damn, dude. You just love trying to be bitchy. You started this thread claiming there "is a STRONG correlation" between per and usg. I posted what the stats say the strength of that correlation is.

You started with a hypothesis about a correlation, and when somebody posts the actual data, you claim you don't see how it relates to the thread?



Um, no. My data / analysis says exactly the opposite. You might want to look into a stats course.



The p-value is ~0, as expected.

I "started" with this. "there appears to be a consistent correlation between a high usage rate and a high PER."

For some reason, you decided to apply that to all players, regardless of PER. Perhaps you just read the first post incorrectly? If so, thanks for the formula, but it's no wonder we aren't on the same page. You supplied data to a question that I never asked.

 

[blazerboy30]

I "started" with this. "there appears to be a consistent correlation between a high usage rate and a high PER."

For some reason, you decided to apply that to all players, regardless of PER. Perhaps you just read the first post incorrectly? If so, thanks for the formula, but it's no wonder we aren't on the same page. You supplied data to a question that I never asked.

Alright, what is a "high PER". If you want to see what the data says, you have to define what "high" means. Let me know and I'll do a new linear regression on players with a per above "x".

 

[PapaG]

Alright, what is a "high PER". If you want to see what the data says, you have to define what "high" means. Let me know and I'll do a new linear regression on players with a per above "x".

PER above 18 would be a decent data set, if you get the time.

Thanks! I think we were on the same page, but speaking different languages because of the other "noise" in the thread. Mea culpa! I was completely misunderstanding what you were showing in terms of data.

I'd love to see the correlation if it exists, between a relatively "high" PER of 17, and the impact of usage rate on that player's PER. The problem with doing this, as I see it, is it still won't account for why one player has a PER of 17 and a usage of 25, while another has a PER of 17 and a usage of 18.

My primary point is that the latter player is more valuable, and would have a higher PER if they had a higher Usage rate. I just don't know the weight of usg and its impact on PER.

 

[blazerboy30]

PER above 18 would be a decent data set, if you get the time.

That works out to be: per = 0.26 * usg + 14.8
r-squared is terrible at 0.27

The "correlation" gets weaker for players at the higher PER.

Thanks! I think we were on the same page, but speaking different languages because of the other "noise" in the thread. Mea culpa! I was completely misunderstanding what you were showing in terms of data.

I'd love to see the correlation if it exists, between a relatively "high" PER of 17, and the impact of usage rate on that player's PER. The problem with doing this, as I see it, is it still won't account for why one player has a PER of 17 and a usage of 25, while another has a PER of 17 and a usage of 18.

That is what the other posters and I have been saying. There is a correlation between PER and usg, that shouldn't be a surprise because they are comprised of many of the same basic stats. The real data shows a correlation, such that on average a player with a higher usg will have a higher PER. But the r-squared value being so bad says that you can't use this analysis to explain the variance in PER on a per-player basis.

 

[PapaG]

That works out to be: per = 0.26 * usg + 14.8
r-squared is terrible at 0.27

The "correlation" gets weaker for players at the higher PER.



That is what the other posters and I have been saying. There is a correlation between PER and usg, that shouldn't be a surprise because they are comprised of many of the same basic stats. The real data shows a correlation, such that on average a player with a higher usg will have a higher PER. But the r-squared value being so bad says that you can't use this analysis to explain the variance in PER on a per-player basis.

I'm not really wondering about the PER variance, per se. I'm wondering if players are being used correctly, among other things, based on a usage basis, considering their PER.

 

[huevonkiller]

This is what he meant:

http://easycalculation.com/statistics/learn-regression.php

http://easycalculation.com/statistics/learn-correlation.php

Just in case anyone here wants to do something similar.

 

[PapaG]

Anyhow, I just emailed John Hollinger a rather lengthy post regarding this, asking for his thoughts on PER, and how he sees usage rate impacting a player's value.

Thought it would be fun to see if he answered it!

 

[PapaG]

This is what he meant:

http://easycalculation.com/statistics/learn-regression.php

http://easycalculation.com/statistics/learn-correlation.php

Just in case anyone here wants to do something similar.

Cool links! Macro, but efficient nonetheless.

 

[blazerboy30]

I'd love to see the correlation if it exists, between a relatively "high" PER of 17, and the impact of usage rate on that player's PER. The problem with doing this, as I see it, is it still won't account for why one player has a PER of 17 and a usage of 25, while another has a PER of 17 and a usage of 18.

My primary point is that the latter player is more valuable, and would have a higher PER if they had a higher Usage rate. I just don't know the weight of usg and its impact on PER.

That is where the correlation coefficient from my regression comes into play. On AVERAGE, a player's PER will increase by 0.5 for every 1pt increase in usg%.

The regression applied to players with a PER above 18 shows that increasing usg% has a smaller effect on increasing PER, at about 0.26 PER increase per 1pt increase in usg%.

 

[Minstrel]

That is where the correlation coefficient from my regression comes into play. On AVERAGE, a player's PER will increase by 0.5 for every 1pt increase in usg%.

The regression applied to players with a PER above 18 shows that increasing usg% has a smaller effect on increasing PER, at about 0.26 PER increase per 1pt increase in usg%.

So there are diminishing returns, which makes sense. If there weren't, you would just put the ball in the hands of your best player every single possession.

There are two competing dynamics here: on the one hand, using more possessions gives you more raw opportunities to do something. However, the more possessions you use, the less you can pick and choose your best chances. As you increase your Usage, each added possession is an increasingly difficult chance, which makes those possessions used less effective.