Is PDSS a good way to analyze the Raptors defense?

[Ripp]

Hi guys,

I figured since a lot of the criticisms I was raising with this methodology didn't seem to be understood, I'd stop arguing in circles and instead show some calculations that follow from this PDSS approach.

I downloaded Boris's PDF file and then put into a spreadsheet, which is now uploaded on Google Docs (see here: https://spreadsheets.google.com/ccc?key ... 0qcB#gid=0)

I took the liberty of adding a new column to his table, what I call "TeamDRat - Drat", column "O". So if your DRat is higher than that of the team's, then you'll get a positive number in green font, and if smaller, a negative number in red font.

As you can see, almost everyone on the team somehow has a better DRat than the team average, except for POB, Rasho, and Jose. Jose's Oliver DRat is in fact 118.30, the worst on the team. As you can see, this by itself seems a bit kooky....almost everybody on the team improves the team defense, yet the defense is somehow terrible? I'll offer no more commentary on this for now, and just let you guys chew on how counterintuitive this result is.

I've broken this post up into four parts:
1) Do PDSS Drtg's match up with what actually happened?
2) Do PDSS Drtg's predict good-performing or bad-performing lineups?
3) Some high-level analysis
4) Concluding Remarks

1) Do PDSS Drtg's match up with what actually happened?

Alright, let's move on. So Row 20 of the sheet has the basketballvalue.com and basketball-reference.com computations for what the team Defensive Rating was this year. The basketballvalue.com simply counts up how many points were scored by opponents of the Raps this year, and then divides by the number of possessions opponents used to score. Obviously, if an opponent scores lots of points on you without using very many possessions, this is very bad...it means that everyone has a highly efficient offense versus you. In other words, your defense is terrible. And as you know, this season we had one of the worst defenses of all time.

Now, usually when you have individual defensive ratings, you'd like to see how they interact to then (A) compute the performance of the entire team and (B) also estimate specific lineups. So let's try to see if we can do task (A) using PDSS.

Alright, so I tried to do this in Rows 27 through 31. Method 1 multiplies each of the individual DRats from PDSS by the number of minutes played, computes the sum, and then divides by the sum of the minutes played by the listed players. In other words, a minutes-weighted average.
Doing this, I get a number (111.7855) that is completely different from the true team DRating from Row 21. In fact, we are off by nearly 1.4 points. In other words, our time-weighted estimate of what the team's defense predicts a better defensive team than expected. Of course, this is obvious if you think about it...Boris's PDSS values indicate that almost everyone on the team is an above-average defender, relative to the team average. So of course once you start taking (time-weighted) averages, you'll get results inconsistent with reality.

Alright, but as Boris said earlier in the previous thread (viewtopic.php?f=32&t=1047850), doing a time-weighted average is NOT the recommended approach to go from individual DRats to team DRats. So let's do a possession-weighted average instead. To compute possessions used for each player (well, something proportional to it), I multiply Dpss% (fraction of defensive possessions used while on the floor) by the minutes played. The result of this possession-weighted average is in Row 28, and is nearly identical with the time-weighting procedure did earlier...we get a team DRat of 111.7203.

So the point of this little calculation is to show you that both time-weighted and possession-weighted estimates should result in very similar team estimates. So if you eyeball a bunch of DRats and all the big-minute guys have GOOD DRats yet the team as a whole has BAD DRats, something is fishy.

Anyway, so we've shown that there is a big gap between the actual team defensive rating and that predicted by this individual PDSS DRat. What happens when we start analyzing individual lineups, say the top 10 lineups most used by the Raps?

2) Do PDSS Drtg's predict good-performing or bad-performing lineups?
Go to column 43, where the Lineup analysis begins. I'll quote from the document:

Next, let's see whether this approach estimates the defensive performance of the
most heavily played Raps lineups. Let's compare with basketballvalue.com (http://basketballvalue.com/teamunits.ph ... 0&team=TOR). The column “Drtg” in blue lists the Drtgs of the corresponding lineups over the season. If the Drtg is Red, then it means that the corresponding lineup defends worse than the team as a whole. If green, then better. Notice that the most used lineup (at ~370 minutes) has defenders who have PDSS Drats better than the team average, yet the end result is a lineup that performs worse than average. Notice that for certain lineups, this strategy of possession-weighting or time-weighting player defensive values cannot mathematically yield anything even close to the correct Drtg for that lineup. For example, lineup #3 (Cackoglu) has a horrifying Drtg of 125.78. Yet, using the possession-weighting scheme that Boris suggested for entire teams, there is absolutely no way to come up with the number 125.78, no matter what normalized-weighting scheme you use.

So, as I've said, even just by eyeballing the ratings for the top 10 most used lineups (H57 to H66), and comparing them to a table of player individual DRats (the light cyan-colored matrix @ I57), we see that there is absolutely no way to go from the individual DRats to the team Drat.

Of course, how big is the discrepancy? We should probably go ahead and calculate the possession-weighted PDSS estimate of the performance of each lineup. To do this, I utilized a table titled "Matrix of PDSS Poss Used", which is right below the cyan-colored matrix @ I57. The results are in the column starting @ N57.

Look closely at the column starting @ N57. Again, all of the numbers are tightly clustered around 110-112. So this technique cannot tell us about certain lineups that were actually used this season and were EXCELLENT defensively, nor can it predict that Cackoglu is a terrible defensive lineup. Of course, if you think about it mathematically, it is clear why this is true...this approach cannot do anything beyond saying a lineup's performance is a weighted average of the players in it. So thus, since the best player's DRat is 107.1 (Pops Mensa Bonsu), and the worse is at 118.3 (Jose Calderon), it has no way of understanding that the Raps had many lineups which performed outside of this range.

Finally, what is the error for the PDSS estimates for each lineup? This is calculated in the cyan-colored column that begins at O56. As I said earlier, it has no clue how to understand either the best Raptors lineups or the worst ones defensively (and again, this makes sense, if you go back to page 25 of the previous thread ((viewtopic.php?f=32&t=1047850)) and read the intuitive explanation I gave.)

3) Some Analysis
As I said on Page 25 of the previous thread (viewtopic.php?f=32&t=1047850), this methodology says a player's DRat = TeamDRat + PlayerOffset (PlayerOffset is computed as a function of Stop% and the rest of the equation Boris posted.) So from there, you are supposed to understand the behavior of lineups and then teams by some sort of averaging procedure (either time-weighted or possession-weighted). It should be clear that if you have any sort of weighted average of a list of say three numbers (N_1, N_2, N_3), it is impossible to get any numbers larger than the biggest number in your list, or smaller than the smallest. So this procedure that Boris outlined has no chance of understanding most of the lineups that teams use during the season.
Of course, someone might just say, "well, it is making mistakes on lineups, but does a good job of predicting the final Team DRtg!" They might say that the error of 1.4 or whatever that I got above is not really that meaningful. But the problem is, we have no way of knowing if the quantity PlayerOffset is correct or not. I could propose an alternate procedure (call it RippRatings) that is almost the exact same as Oliver's Drtg, but instead sets Playeroffset to a uniformly random variable between say -.1 and +.1. RippRatings are of course stupid and offer no insight into player performance, will fail to describe individual lineups (just as Oliver's ratings do), but will do even BETTER when it comes time to compute the Team Drtg (since I'm just averaging several random variables between -.1 and +.1).
So this is the flaw of the method...it doesn't understand lineups, and the fact that it produces numbers close to the average team rating is meaningless, since this useless RippRatings procedure I did does the same exact thing, while offering no insight into who is good and who isn't.

4) Concluding Remarks
Feel free to download this sheet and play around with it yourself. And if you find any errors I made, again bring them to my attention.

Finally, just a comment...I know it might seem that I think this work is terrible, but this isn't true. I think there are some interesting ideas here. Like I said on page 25 of the other thread, the variable TeamDRat - OliverDrat is very interesting to me, and has the same feel of ON/OFF or APM style quantities. I also think Stop% is great, though I don't necessarily agree with this specific formula used here. But the idea itself is great, since you might actually be able to build some models involving variants of Stop% that do a good job of measuring a player's defensive impact, and then go from there to understanding lineup performance, and finally team performance.

So in short, the ideas here are useful, even though the exact formulation and "holy grail" approach that Boris takes less useful. In particular, you should be suspicious of his Lake Woebegonism, which seems to argue that almost everyone on the Raps is above average defensively, and that Bargs is one of the best defenders on the team. Next thing we will hear from him is that the Raptors also have always been at war with Eurasia...

[Ripp]

OK, I realize the above block of text is pretty intimidating. So to summarize:

• a) PDSS is a technique that cannot figure out how good different Raps lineups are defensively.
• b) It also cannot even figure out the Raptors Drtg for the season, despite this being a piece of information I provide to it.
• c) I illustrated how this flaw raises serious questions about how useful it is by giving a new rating system that is similar to PDSS, also cannot figure out how good lineups are, but does even better in figuring out the Raps Drtg.
• d) However, this rating system, despite doing a better job than PDSS in figuring out the Raps Drtg is clearly useless, since it just assigns ratings to players randomly...thus it actually doesn't give us any insight into how good the defenders on the team are.

[supersub15]

I don't have time now to go over the details of PDSS and of what you've done, but have you considered that Boris logged 8% of the plays under the "team" category, which may (or may not) account for the discrepancy?

[Ripp]

Intuitively, 8% of plays is not going to be the difference for an algorithm that is broken, or doesn't measure the right thing...unless we have strong evidence indicating that these 8% of plays are very special in some way (i.e., there is bias in the way they were chosen.)

Still, those 8% of plays that were not incorporated...I tried a time-weighted estimation procedure that treats "Team" as as separate individual. This is the number in Method 2, on line 27 of the above spreadsheet (the value 112.3554). Again, this uses time-weighting rather than possession-weighting. Just now on my computer, I added a new column to redo "Method 3" (possession-weighted), but again incorporating it as a new player. I get the value 111.998.

I don't want to modify the original spreadsheet I made to take this new variable into account, otherwise I'll have to edit my post accordingly (and change my references from NBlah to N(Blah+1) r whatever.)
I've uploaded it to a separate spreadsheet here called Boris2:
https://spreadsheets.google.com/ccc?key ... ybII#gid=0

The two new weighted ratings that take time and usage into account are on lines 27 and 29, respectively.

Again, for those of you trying to read and understand the original post, use the spreadsheet linked there, not this one...otherwise the instructions I give in the original post will be off.

[Marlowe]

Hey Ripp,

Intriguing stuff. I am on go train as I write this and haven't had a good chance to look at your chart. I wanted to ask a question . Do you think it is possibe to put weighting or adjust by each players position. Perhaps Calderons position has more an impact on defence in our team than any other? Just a thought.

[elmer_yuck]

>>As you can see, almost everyone on the team somehow has a better DRat than the team average,
>>except for POB, Rasho, and Jose. Jose's Oliver DRat is in fact 118.30, the worst on the team.
>>As you can see, this by itself seems a bit kooky....almost everybody on the team improves the
>>team defense, yet the defense is somehow terrible?

That pretty much kills this stat.

[Ripp]

Hey Ripp,

Intriguing stuff. I am on go train as I write this and haven't had a good chance to look at your chart. I wanted to ask a question . Do you think it is possibe to put weighting or adjust by each players position. Perhaps Calderons position has more an impact on defence in our team than any other? Just a thought.

This is a very good point you raise. Someone might argue that, "Hey, maybe PDSS itself is a good system, but you aren't just weighting player ratings properly to understand lineup performance! For example, maybe a PG should be worth 40%, and all other players total 60%.

However, take a look at the Lineup on line 59 of the spreadsheet:

Code: Select all

Lineup Calderon-Jack-Turkoglu-Bargnani-Bosh
PDSS DRat 118.3000   112.8000   112.3000   110.0000   110.0000
Actual Drtg for the lineup 125.78


Again, when I say "Actual Drtg for the lineup", I mean the number of points scored on a lineup divided by number of possessions, the same sense that basketballvalue.com uses. Not this PDSS Individual Drtg stuff. So it was one of the worst lineups defensively, of the 10 most played.

So even if you use a weighting scheme that assigns all of the mass to Calderon (i.e., he gets a weighting of 1 and everyone else gets 0), you are still very far away from the actual performance of 125.78. In fact, the maximum you can get is just Jose's value...118.30. So in other words, even if you assign ALL the blame onto Jose, it still doesn't explain how bad this lineup was defensively.

Even worse...remember, I'm not allowed to change this weighting scheme from lineup to lineup. So in layman's terms, if I am putting a lot of blame on Jose, then it had better damn sure turn out that lineups that DON'T involve Jose are VERY GOOD defensively.

But this is not the case! So many of those top 10 lineups are bad. So Jose cannot be the only one at fault. Translating back into math terms, a weighting scheme that assigns all of the mass to Jose is almost certainly not does not lead to an accurate model of what actually happened.

So to summarize, absolutely no amount of re-jiggering of the weights will allow the PDSS numbers of Boris to explain the actual lineup performance. Given this, there is no reason to believe in the validity of this method....why should I believe a method that doesn't explain what actually happened on the floor (i.e., lineup performance)?

[BorisDK1]

Alright, let's move on. So Row 20 of the sheet has the basketballvalue.com and basketball-reference.com computations for what the team Defensive Rating was this year. The basketballvalue.com simply counts up how many points were scored by opponents of the Raps this year, and then divides by the number of possessions opponents used to score. Obviously, if an opponent scores lots of points on you without using very many possessions, this is very bad...it means that everyone has a highly efficient offense versus you. In other words, your defense is terrible. And as you know, this season we had one of the worst defenses of all time.

Now, usually when you have individual defensive ratings, you'd like to see how they interact to then (A) compute the performance of the entire team and (B) also estimate specific lineups. So let's try to see if we can do task (A) using PDSS.

Alright, so I tried to do this in Rows 27 through 31. Method 1 multiplies each of the individual DRats from PDSS by the number of minutes played, computes the sum, and then divides by the sum of the minutes played by the listed players. In other words, a minutes-weighted average.

Well, that's a problem because not every player faces the same number of possessions. "Team" (.080 DPoss%) faces fewer of the opponents' possessions while on the floor than, say, Amir Johnson (.228 DPoss%). (Well, "team" is actually always on the floor.) A better way of doing this is summing total individual defensive possessions and dividing by points allowed (estimating for the value of a FGA, because I didn't separate 3FGA this year). I really think you missed that, even though it was pointed out numerous times. This is not a system of merely counting how many points the opponents scored while a player was on the floor, of course. Not every player faces the same amount of possessions.

Oh, a brief note on player minutes - I had trouble getting Excel to handle min:sec in calculating a lot of the things that need to get calculated with player minutes, so I rounded each game by .25s. Thus, 15:07 of playing time got rounded to 15, 15:8 got rounded to 15.25, etc.

Doing this, I get a number (111.7855) that is completely different from the true team DRating from Row 21.

Keep in mind that the formulas for the DRat include only those games in which the player actually appeared.

In fact, we are off by nearly 1.4 points. In other words, our time-weighted estimate of what the team's defense predicts a better defensive team than expected. Of course, this is obvious if you think about it...Boris's PDSS values indicate that almost everyone on the team is an above-average defender, relative to the team average. So of course once you start taking (time-weighted) averages, you'll get results inconsistent with reality.

Alright, but as Boris said earlier in the previous thread (viewtopic.php?f=32&t=1047850), doing a time-weighted average is NOT the recommended approach to go from individual DRats to team DRats. So let's do a possession-weighted average instead. To compute possessions used for each player (well, something proportional to it), I multiply Dpss% (fraction of defensive possessions used while on the floor) by the minutes played. The result of this possession-weighted average is in Row 28, and is nearly identical with the time-weighting procedure did earlier...we get a team DRat of 111.7203.

DPoss% multiplied isn't the way to go, since as explained I'm only accounting for games in which those individuals actually appeared. DPoss% for one player is going to be calculated from a different set as for another. Summing all the individual possessions gives you a total amount of individual possessions of 7580, awfully darn close to the estimated number of possessions of 7670 - which can probably be explained by effects of rounding, etc.

So the point of this little calculation is to show you that both time-weighted and possession-weighted estimates should result in very similar team estimates. So if you eyeball a bunch of DRats and all the big-minute guys have GOOD DRats yet the team as a whole has BAD DRats, something is fishy.

Ummm...the best DRat on the team of all the rotation players was 109.5 (Amir Johnson's), which by my count is still almost 2 points / 100 possessions worse than league average. How is that "good", exactly?

Anyway, so we've shown that there is a big gap between the actual team defensive rating and that predicted by this individual PDSS DRat. What happens when we start analyzing individual lineups, say the top 10 lineups most used by the Raps?

Not when you calculate using the data properly - and including "team", as supersub15 pointed out and which you so badly missed.

I'm not going to respond to the lineup-matchups data because the latter doesn't handle direct performance at all.

As I said on Page 25 of the previous thread (viewtopic.php?f=32&t=1047850), this methodology says a player's DRat = TeamDRat + PlayerOffset (PlayerOffset is computed as a function of Stop% and the rest of the equation Boris posted.)

Man, another basic mistake of fact.

DPoss% is the "player offset" you mention, and is not calculated from the stop% at all. I've pointed out that - what, a good half a dozen times? I don't know what I have to do to get that across. Do I have to hire a skywriter? Burn it into your lawn with diesel and fertilizer? Paint it in blood on your wall? Just let me know, I'm willing to go the extra mile for the sake of education, here.

Of course, someone might just say, "well, it is making mistakes on lineups, but does a good job of predicting the final Team DRtg!" They might say that the error of 1.4 or whatever that I got above is not really that meaningful. But the problem is, we have no way of knowing if the quantity PlayerOffset is correct or not.

Oh, my goodness.

DPoss% = (Individual Defensive Possessions x (Team Minutes / 5) ) / (Minutes Played x Team Possessions). Yeah, really tough to figure out there.

Finally, just a comment...I know it might seem that I think this work is terrible, but this isn't true. I think there are some interesting ideas here. Like I said on page 25 of the other thread, the variable TeamDRat - OliverDrat is very interesting to me, and has the same feel of ON/OFF or APM style quantities. I also think Stop% is great, though I don't necessarily agree with this specific formula used here. But the idea itself is great, since you might actually be able to build some models involving variants of Stop% that do a good job of measuring a player's defensive impact, and then go from there to understanding lineup performance, and finally team performance.

So in short, the ideas here are useful, even though the exact formulation and "holy grail" approach that Boris takes less useful. In particular, you should be suspicious of his Lake Woebegonism, which seems to argue that almost everyone on the Raps is above average defensively, and that Bargs is one of the best defenders on the team. Next thing we will hear from him is that the Raptors also have always been at war with Eurasia...

I do think PDSS data is far more reliable to assess a team's defense with than, say, indirect approaches. I'm not opposed to looking at the indirect data, as long as we're using it in the limited manner that in which it's proper to b e used. Individual play is being evaluated, you can see the guys active defensively vs. not, you can see the guys doing a good job stopping guys or not and you can see the ultimate outcome. And, contrary to your assertion, not one player on the Raptors was an above-average defender this year with the DRat. Not a single one. Yes, some guys are more responsible for allowing scores less efficiently on the team than others - what a radical concept!

Ultimately, what you've done has just been very poor in fact-checking and trying to take in what others are saying. And that really diminishes any value your critique might have, especially when even supersub15 points out something you've missed and you pooh-pooh it.

8% of plays on every single team minute is a big number, more than 20% of the plays on less than a hundred minutes of individual playing time, for example.

[Ripp]

Jesus Christ man....step back away from the math a second, use your brain and THINK.

Your calculation says JOSE alone is the only major-minute guy who is below average defensively (Jose, Rasho, and POB, to list them all.) So if I have a lineup that doesn't feature ANY of these players, how the hell can the resulting lineup still be BELOW average? I've removed all the below average guys from my lineup, and still end up with a below-average lineup?

In math terms, I have a list of the following numbers:
1, 5, 10, 30
The average of these numbers is 11.5. If I then remove the number 30 from the list, the average of the remaining numbers has to be smaller than 11.5. Do you understand this, or not?

[OvertimeNO]

Sorry to jump in, but I just want to let you guys (mainly Boris, Ripp, and SS) know I appreciate this discussion.

At some point it would be really interesting if you guys could test the predictive aspects of your preferred statistical analysis methods, once our final opening day lineup is set.

[ranger001]

What is the formual for individual drat and team drat? Maybe the two cannot be correlated by the formula you use.

[OvertimeNO]

Oh, and to add my own completely uneducated two cents:

I propose that "team defense" as it's being considered here is an emergent property that would be extremely difficult to account for, both when analyzing individual performance based on team ratings, and calculating team performance based on individual ratings. Taken to the most extreme, it's like analyzing a thought to directly extrapolate what an individual brain cell is doing, and vice versa. Obviously the problem here isn't as complex as the brain/mind issue, but ten unique and autonomous beings running around in an enclosed area still seems to be a complex enough interaction that you have to insert an extra level of analysis to account for the overall result.

[Ripp]

Well, that's a problem because not every player faces the same number of possessions. "Team" (.080 DPoss%) faces fewer of the opponents' possessions while on the floor than, say, Amir Johnson (.228 DPoss%). (Well, "team" is actually always on the floor.) A better way of doing this is summing total individual defensive possessions and dividing by points allowed (estimating for the value of a FGA, because I didn't separate 3FGA this year). I really think you missed that, even though it was pointed out numerous times. This is not a system of merely counting how many points the opponents scored while a player was on the floor, of course. Not every player faces the same amount of possessions.

I invite you to do this yourself, and report your findings. The spreadsheet is freely available online now, so you can go ahead and use any weighting scheme you wish. As I stated, it is impossible for any weighting scheme to explain the behavior of many of the top 10 most frequently played lineups. But you are welcome to try...if you can find one, I'll be quite impressed.

Ummm...the best DRat on the team of all the rotation players was 109.5 (Amir Johnson's), which by my count is still almost 2 points / 100 possessions worse than league average. How is that "good", exactly?

If it isn't clear, I mean good and bad relative to the team's DRTG for the entire season.

Not when you calculate using the data properly - and including "team", as supersub15 pointed out and which you so badly missed.

Did you not read my post and the new attached spreadsheet which DOES include team, and again in which it doesn't equal the actual Team Drtg? Did you miss my comment in which for some of those lineups is it mathematically impossible for ANY weighting scheme to predict the lineup performance?

I'm not going to respond to the lineup-matchups data because the latter doesn't handle direct performance at all.

Heh, OK. In other words, this approach is useless for predicting how lineups will actually do.

Oh, my goodness.

DPoss% = (Individual Defensive Possessions x (Team Minutes / 5) ) / (Minutes Played x Team Possessions). Yeah, really tough to figure out there.

Go to page 25 of the previous thread...PlayerOffset I'm defining to be DRat - TeamDRat. Given this definition, you should reread my criticism and hopefully understand it.

[dacrusha]

Do all of these statistical semantics really matter? Bottom line is both Bargs and Jose suck at defense.

The only thing really worth arguing about is which position (PG or C) do you think has more impact on the overall strength/weakness of a team's defense.

[Ripp]

Do all of these statistical semantics really matter? Bottom line is both Bargs and Jose suck at defense.

The only thing really worth arguing about is which position (PG or C) do you think has more impact on the overall strength/weakness of a team's defense.

You are 100% correct. But I just wanted to make sure that people did not come out the previous thread thinking that Bargnani is indeed the 3rd best defender on the team, as this PDSS DRat indicates. Hopefully I've done a good job explaining why the methodology here is flawed, and should not be used to draw conclusions like this. Heck, maybe in reality Bargs is the 3rd best defender on the team....but you'll have to show it a different way than by using PDSS DRat.

[BorisDK1]

Jesus Christ man....step back away from the math a second, use your brain and THINK.

Your calculation says JOSE alone is the only major-minute guy who is below average defensively (Jose, Rasho, and POB, to list them all.) So if I have a lineup that doesn't feature ANY of these players, how the hell can the resulting lineup still be BELOW average? I've removed all the below average guys from my lineup, and still end up with a below-average lineup?

In math terms, I have a list of the following numbers:
1, 5, 10, 30
The average of these numbers is 11.5. If I then remove the number 30 from the list, the average of the remaining numbers has to be smaller than 11.5. Do you understand this, or not?

So what we have here is anger over results that you don't like, therefore you question the method (without understanding it particularly well). So I guess we're not supposed to pay attention to your numerous mistakes of fact, we'll sweep all those under the rug - we've got some hardcore venting that something has come along to challenge statboy assumptions.

I'm going to get offensive here but it's not necessarily directed at you personally for the most part. I get really tired of attitude pretty quickly: I have no truck for it. I hate the attitude that comes from a certain segment that says, "who needs stats? I get everything I need with my eyes" and the other segment that says, "I've got my spreadsheets and some methods I really like and this is gospel and anything that doesn't agree goes out the window". And to be perfectly honest, you are acting very much the latter in this case. I'm not passing judgment on you or your basketball acumen, but in this discussion your conduct is very much the latter.

[Reignman]

Jesus Christ man....step back away from the math a second, use your brain and THINK.

Your calculation says JOSE alone is the only major-minute guy who is below average defensively (Jose, Rasho, and POB, to list them all.) So if I have a lineup that doesn't feature ANY of these players, how the hell can the resulting lineup still be BELOW average? I've removed all the below average guys from my lineup, and still end up with a below-average lineup?

In math terms, I have a list of the following numbers:
1, 5, 10, 30
The average of these numbers is 11.5. If I then remove the number 30 from the list, the average of the remaining numbers has to be smaller than 11.5. Do you understand this, or not?

So what we have here is anger over results that you don't like, therefore you question the method (without understanding it particularly well). So I guess we're not supposed to pay attention to your numerous mistakes of fact, we'll sweep all those under the rug - we've got some hardcore venting that something has come along to challenge statboy assumptions.

I'm going to get offensive here but it's not necessarily directed at you personally for the most part. I get really tired of attitude pretty quickly: I have no truck for it. I hate the attitude that comes from a certain segment that says, "who needs stats? I get everything I need with my eyes" and the other segment that says, "I've got my spreadsheets and some methods I really like and this is gospel and anything that doesn't agree goes out the window". And to be perfectly honest, you are acting very much the latter in this case. I'm not passing judgment on you or your basketball acumen, but in this discussion your conduct is very much the latter.

Attitude aside, the question is extremely valid. If individual PDSS shows Jose as the only rotation player that's below average then why does the overall defense still stink more often than not when he's not in the lineup? Especially considering PDSS shows the other players as above average individual defenders.

[Ripp]

Boris: I apologize if I came across as rude, it was not my intention. But I'd still like some answers to those questions I'm asking about this technique.

[BorisDK1]

I invite you to do this yourself, and report your findings. The spreadsheet is freely available online now, so you can go ahead and use any weighting scheme you wish. As I stated, it is impossible for any weighting scheme to explain the behavior of many of the top 10 most frequently played lineups. But you are welcome to try...if you can find one, I'll be quite impressed.

As stated before, I have no problem looking at lineup data, but the attempt here to use that data to turn that into a stat that analyzes individual defense is fallacious.

If it isn't clear, I mean good and bad relative to the team's DRTG for the entire season.

And that's not really all that important, when 8% of the team's possessions aren't being covered by anybody (and giving up a lot of scores).

Heh, OK. In other words, this approach is useless for predicting how lineups will actually do.

Not at all. Any coach can look at the performance of his players and realize that putting certain defenders on the floor in certain circumstances is going to be a bad idea.

Go to page 25 of the previous thread...PlayerOffset I'm defining to be DRat - TeamDRat. Given this definition, you should reread my criticism and hopefully understand it.

I understand your criticism, and I've tried to interact with it fairly (I wish this method would get the same treatment from you!). But it's funny that when you sum up all the individual defensive possessions, and you estimate the value of FGA and sum that all up from PDSS data, you get something that looks an awful hell of a lot like the team's actual DRat of 113.2. It gives you 113.9, and that accuracy will improve next year when I specifically track 3FGA so I'm not left guessing about the value of FGA.

So the "smell test" that you kept referring to, doesn't seem quite so outrageous. It seems to more or less work, far more accurately than you're willing to give it credit for.

Your objection to the variance still sounds an awful lot like, "the results I see there don't agree with what I think, so I'll object."

[BorisDK1]

Attitude aside, the question is extremely valid. If individual PDSS shows Jose as the only rotation player that's below average then why does the overall defense still stink more often than not when he's not in the lineup? Especially considering PDSS shows the other players as above average individual defenders.

The answer is, "team" (uncontested action) isn't really being accounted for. 8% of the team's possessions went unaccountable to any specific player this year, and that is going to account for how things look for individuals.