Manu Ginobili and playing at a starter disadvantage

[lessthanjake]

Be interested to see Harden's numbers here, especially during his OKC days. 2011-2012 especially, those Harden bench lineups were pretty absurd. Curious to see if that translated to his time on the Rockets or Nets.

Surprisingly, Harden has generally not done well at all at a starter disadvantage. Below are his numbers, filtering out minutes with Durant, Westbrook, Howard, CP3, Kyrie, Embiid, Kawhi, and Paul George. Like with Ginobili, I am treating Harden as a starter for these purposes in the years he was a bench player.

James Harden at a Starter Disadvantage - Without Durant, Westbrook, Howard, CP3, Kyrie, Embiid, Kawhi, and Paul George (always treating Harden as a starter)

2010: -46 in 176 possessions
2011: +8 in 302 possessions
2012: -8 in 240 possessions
2013: +32 in 1017 possessions
2014: -66 in 475 possessions
2015: -38 in 1393 possessions
2016: -66 in 938 possessions
2017: -8 in 1525 possessions
2018: +78 in 1177 possessions
2019: -101 in 1244 possessions
2020: +21 in 595 possessions
2021: +13 in 446 possessions
2022: -3 in 476 possessions
2023: -49 in 410 possessions
2024: -8 in 349 possessions

If we look at 2012-2024, Harden has a net rating of -1.97 at a starter disadvantage without any other star on the court with him.

I’ll note that if we don’t do the star filtering, Harden was +13 in 555 possessions (+2.34 net rating) in that 2012 season you were most curious about.

[trex_8063]

How can you consider Kevin Love a star from 2016-2018 yet not consider Brook Lopez, Klay Thompson or Andre Iguodala stars?

These numbers also greatly benefit the teams who are super deep 8-12 in the rotation, while hurting teams who have more higher-salaried players near the top since those teams will inevitably have less depth and more minimum-level players.

It also benefits teams who went deep into the luxury tax to retain lesser bench pieces [See Golden State] and hurt teams who avoided the Tax or 2nd Apron.

You know why

Reminds me of another poster who came up with WOWY “adjustments” for LeBron that included 2011 Shaq but of course not 2015 Varejao who got injured as well as expansion caveats for early ‘70s Kareem but not ‘90s Jordan.

As an aside, I don't see that he stuck the "star" label with any of these guys; he took out who he took out and fed us the numbers. If you'd like to see minus additional supporting cast players [for Giannis or Steph, or whomever], you can run the numbers yourself and share it as a counterpoint.

Jabbing like this after a big share of info [that I imagine took some time to compile] is not helpful (and probably harmful). Please don't do it.

[Colbinii]

I wonder what advantage someone playing with, say, 2015-2017 Andre Iguodala off the bench has, especially when it is by far the best line-up for the Warriors and he closes the game with 4 starters.

This data is cool, but what exactly does it tell us?

1) We aren't controlling for quality of bench player(s)
2) We aren't controlling for time spent with each bench player(s)
3) We aren't controlling for time spent with starter(s)

For example, Mario Chalmers and Chris Bosh are both starters for the 2011 Heat. Nobody needs to ask which 4v5 line-up is better, the one with Chris Bosh as one of the 4 starters or the one with Mario Chalmers.

Furtheremore, the 2007 Spurs [for example], started two centers who played < 20 MPG. Francisco Elson and Fabrico Oberto. Sub in Manu for either one and either Brent Barry or Michael Finley and boom the Spurs are closing 4v5 in terms of Starters for each team. Is this really impressive? Because the data here is telling us it should be impressive.

[lessthanjake]

I wonder what advantage someone playing with, say, 2015-2017 Andre Iguodala off the bench has, especially when it is by far the best line-up for the Warriors and he closes the game with 4 starters.

We can check! From 2015-2017, the Warriors at a starter disadvantage with Steph and Iguodala on had a +14.77 net rating, and in those same years at a starter disadvantage with Steph on and Iguodala off, the Warriors had a +11.24 net rating. So, as we’d expect, Iguodala helps, but the numbers were great regardless. And Steph’s player-filtered data here is actually slightly *better* in the 2020s where Iguodala was either not on the Warriors or was an essentially irrelevant player. So obviously Iguodala was a good player, but I don’t think he’s driving the numbers for Steph.

This data is cool, but what exactly does it tell us?

1) We aren't controlling for quality of bench player(s)
2) We aren't controlling for time spent with each bench player(s)
3) We aren't controlling for time spent with starter(s)

These are fair points. But I think I already made relevant caveats regarding these issues in my OP, when I talked about not controlling for the quality of bench players, and also mentioning that this gets at stuff already accounted for in RAPM (which does control for the #2 and #3 things you mentioned). In an ideal world, we might run some sort of starter-disadvantage-specific RAPM such that we are controlling for these things you mention, but I don’t have the time or data to do that.

For example, Mario Chalmers and Chris Bosh are both starters for the 2011 Heat. Nobody needs to ask which 4v5 line-up is better, the one with Chris Bosh as one of the 4 starters or the one with Mario Chalmers.

Furtheremore, the 2007 Spurs [for example], started two centers who played < 20 MPG. Francisco Elson and Fabrico Oberto. Sub in Manu for either one, and boom the Spurs are closing 4v5 in terms of Starters for each team. Is this really impressive? Because the data here is telling us it should be impressive.

I don’t think these are really fair examples of issues with the data I’ve presented, since you raise issues I already explicitly made adjustments for. For the first one, this exact sort of thing is why I presented a version of the data that filtered out someone like Chris Bosh. And for the second one, that’s exactly why I presented versions of Ginobili’s data that treats him as a starter. In those versions of the data, the scenario you describe there would *not* be part of the data at all, because I’d be treating Ginobili as a starter for these purposes (meaning that scenario would be treated as 5v5 in terms of starters).

[Colbinii]

I don’t think these are really fair examples of issues with the data I’ve presented, since you raise issues I already explicitly made adjustments for. For the first one, this exact sort of thing is why I presented a version of the data that filtered out someone like Chris Bosh. And for the second one, that’s exactly why I presented versions of Ginobili’s data that treats him as a starter. In those versions of the data, the scenario you describe there would *not* be part of the data at all, because I’d be treating Ginobili as a starter for these purposes (meaning that scenario would be treated as 5v5 in terms of starters).

My point is both Finley AND Barry were better than the Starting Center for the 2007 Spurs team. So, when you have Barry/Finley in and then Manu/Parker/Duncan/Bowen, that's really the best line-up and is going to skew "Bench data" if we want to use it holistically.

[lessthanjake]

I don’t think these are really fair examples of issues with the data I’ve presented, since you raise issues I already explicitly made adjustments for. For the first one, this exact sort of thing is why I presented a version of the data that filtered out someone like Chris Bosh. And for the second one, that’s exactly why I presented versions of Ginobili’s data that treats him as a starter. In those versions of the data, the scenario you describe there would *not* be part of the data at all, because I’d be treating Ginobili as a starter for these purposes (meaning that scenario would be treated as 5v5 in terms of starters).

My point is both Finley AND Barry were better than the Starting Center for the 2007 Spurs team. So, when you have Barry/Finley in and then Manu/Parker/Duncan/Bowen, that's really the best line-up and is going to skew "Bench data" if we want to use it holistically.

The lineups you are talking about are essentially irrelevant for purposes of starter-disadvantage analysis. From 2005-2010 (the time period where at least one of Finley or Barry were on the team), the Spurs played a grand total of 84 possessions at a starter disadvantage with Duncan/Parker/Bowen/Manu all on and one of Finley or Barry on as well (treating Ginobili as a starter for these purposes). And they went +2 in those 84 possessions. It’s basically irrelevant and not even helpful to Ginobili anyways.

And it’s the type of thing that could be nitpicked about most teams. For instance, LeBron’s data on the Heat is surely sweeping in minutes where Joel Anthony or Udonis Haslem (who were starters) are off and Shane Battier (a better player, who wasn’t a starter) was on. And the same sort of thing could be true of opponents, who might have bench players that are better than their starters, leading to their minutes with a starter advantage not actually being as strong as we’d think. It’s not wrong to point out that starter state isn’t a perfect proxy for the relative quality of the players on the court. The fact that it’s not a perfect proxy is inherently a blind spot in the data I’ve presented. But I don’t think this is at all specific to the Spurs.

[jalengreen]

Very cool data, haven't seen this sort of thing analyzed before so I appreciate you sharing.

This sort of thing is definitely tricky because of all the potential adjustments you can make; the Klay thing, as has been pointed out, and the fact that removing him makes Steph's numbers look far better (and decreases the sample size further to like, half a season's worth of possessions I think?). Lot of different ways to try to adjust things and get very different values. Of course it seems like no matter how you slice it, Manu's numbers are ridiculously impressive.

I calculated Steph's numbers with Klay off the court, but those have already been posted. Also LeBron 2009-13: net rating with a starter disadvantage and sans Wade/Bosh is +10.99.

Yeah, agreed. One can make lots of different filtering decisions and it can swing the numbers (as we saw with the Klay thing). It takes some time to run these, so the only numbers I ran for any of these guys are exactly what I’ve reported out. So I don’t really know what further filtering would do in specific cases, but I suspect it could change things (even if only just due to randomness). What I ran was just the filtering I thought made the most sense, without knowing beforehand how those filtering decisions would affect the data.

And yeah, I actually think that LeBron’s 2009-2013 numbers being really great here in a sense tends to validate that the method is measuring something meaningful, since it does show LeBron’s consensus best years as being the ones where he looks best in this analysis (even down to 2011 being a down year in that timeframe). I take that as an encouraging sign about the analysis.

That said, I’d note again that lumping in 2009 and 2010 with the players-filtered-out version is a bit iffy, since filtering out stars inherently means more of these minutes are with bench or deep bench guys, and not filtering them out is very likely going to leave a lot larger a proportion of the minutes being with the other starters (but just in starter states of 4v5, 3v5, and 3v4). And, without filtering out anyone, that’d include often having the next best guys on the team on (which obviously isn’t the case with years where star teammates are filtered out). So it’s a bit comparing apples and oranges to compare filtered and non-filtered years. But, as I said in my OP, there’s not an obvious person to filter out in those years, so I don’t really have a better solution. My inclination is to think that any filtering in 2009 and 2010 probably wouldn’t change the bottom-line number for 2009-2013 much anyways, since it’d probably make the 2009 and 2010 numbers less good but also make those years a smaller part of the sample, leaving the super high 2012 and 2013 numbers as a larger part of the data. For instance, if we filtered out Mo Williams and Varejao, we get +24 in 316 possessions in 2009, and +23 in 233 possessions in 2010. Those unsurprisingly look less good than the non-filtered numbers for those years do, but the overall 2009-2013 time period ends up being +10.34 anyways, which is barely lower than what you listed above and is still outrageous.

I will note that Manu has timeframes that are pretty outrageous too. For instance, his 2012-2016 span with stars filtered is a +13.75 net rating.

There's no way to do this cleanly once you get into those adjustments, which is kinda what this shows. Yeah, if you don't filter out anyone for the first stint Cavaliers then you're not filtering out LeBron's best teammates.

But comparing the that first stint Cavs lineups to the disadvantaged Steph lineups shows that you're obviously not leaving them with the same level of non-starting teammates

- The filtered LeBron first stint Cavs lineup with the most possessions in a disadvantaged starter state is LeBron, Gibson, Hickson (20-years-old), Szczerbiak (out of league next year), Pavlovic. +8 in 30 possessions

- Next up is LeBron, Gibson, Hickson, Kinsey, Szczerbiak. Don't even remember hearing of Tarence Kinsey lol, apparently he was also out of the league after the 2009 season (tore it up in Europe though). +9 in 26 possessions

While filtered Steph lineups could be something like:

- Steph/Poole/Klay/Wiggins/Looney (+10 in 43 possessions)
- Steph/CP3/Klay/Kuminga/Looney (+19 in 20 possessions)
- Steph/GP2/Bjelica/OPJ/Iguodala (+11 in 46 possessions)

A problem with RAPM is interpretability. That's why being able to say "the Spurs had a +6.00 net rating from 2003-2017, at a starter disadvantage and Ginobili on and Duncan, Parker, Robinson, and 2014-onwards Kawhi all off" is pretty cool, and certainly an interesting tidbit. But there are concerns with a comparative approach, and I think LeBron/Curry is an extreme example that illustrates that (first stint Cavs dearth of talent vs Strength in Numbers). Like you said, comparing filtered to non filtered years is apples to oranges, but so is the bench players of one team compared to the bench players of another. That's where RAPM comes in and helps. Not to say you did anything wrong, just something to consider.

[Colbinii]

And it’s the type of thing that could be nitpicked about most teams. For instance, LeBron’s data on the Heat is surely sweeping in minutes where Joel Anthony or Udonis Haslem (who were starters) are off and Shane Battier (a better player, who wasn’t a starter) was on. And the same sort of thing could be true of opponents, who might have bench players that are better than their starters, leading to their minutes with a starter advantage not actually being as strong as we’d think. It’s not wrong to point out that starter state isn’t a perfect proxy for the relative quality of the players on the court. The fact that it’s not a perfect proxy is inherently a blind spot in the data I’ve presented. But I don’t think this is at all specific to the Spurs.

It isn't specific to the Spurs, THAT IS MY POINT. My point is the samples are small, the difference between the quality of players like "Shouldn't be playing any minutes" and "Good 6th man" is massively large. None of your data accounts for the massive differences in quality of benches.

You can't just say "Well everyone has asterisks so let's just ignore them" because that isn't a logical way to use this data.

[Colbinii]

2011-2014 Miami Heatles

Spoiler:

Regular Season

2011-2014 LeBron James
6125 Possessions
+9.29 Net +/-

2011-2014 Dwyane Wade
3791 Possessions
+6.74 Net +/-

2011-2014 Chris Bosh
4180 Possessions
+10.30 Net +/-

2011-2014 LeBron James (No Wade, No Bosh)
2386 Possessions
+4.23 Net +/-

2011-2014 Dwyane Wade (No LeBron, No Bosh)
735 Possessions
-9.20 Net +/-

2011-2014 Chris Bosh (No LeBron, No Wade)
607 Possessions
+10.17 Net +/-

Post-Season
2011-2014 LeBron James
2165 Possessions
+5.86 Net +/-

2011-2014 Dwyane Wade
1513 Possessions
+1.35 Net +/-

2011-2014 Chris Bosh
1149 Possessions
+5.75 Net +/-

2011-2014 Chris Bosh (No LeBron, No Wade)
26 Possessions
-27.18 Net +/-

2011-2014 Dwyane Wade (No LeBron, No Bosh)
135 Possessions
-8.62 Net +/-

2011-2014 LeBron James (No Wade, No Bosh)
696 Possessions
+10.71 Net +/-

Now this is some fun data to look at. In the Regular Season, we see that LeBron and Bosh were head-and-shoulders above Wade in these line-ups with more bench players, but where is really becomes drastic is just how many more possessions LeBron played without both Bosh and Wade. LeBron played nearly twice as many possessions without both Bosh and Wade did than both Wade and Bosh played COMBINED without the other two. It's quite clear that Spoelstra relied heavily on the LeBron and No Bosh + No Wade line-ups throughout their time, especially in the 2012-2014 seasons as Wade's body started to fall apart and Bosh missed some games.

Jumping to the post-season, this is the real deal here. Wade was anything but a Star during the totality of the 4 post-season runs, and LeBron's ability to muster a +10.71 Net Rtg WITHOUT Wade/Bosh in nearly 700 possessions is absolute insanity. One, it is insane because 700 Possessions without your 2nd AND 3rd best players, over a 4-year playoff run, is unheard of. Two, that's GOAT level post-season Net-Rtg all done without anything close to an all-star in the line-up.

For reference, Curry, in 5 years [2015-2019], only played 242 Possessions in the post-season WITHOUT Klay AND Draymond. LeBron nearly tripled that amount of possessions in 1 less year

The other player who I want to look at is Chris Paul, who comes out looking horrible with the Clippers.
2014-2017 Chris Paul
2341 Possessions
+0.36 Net +/-

[Colbinii]

I wonder what advantage someone playing with, say, 2015-2017 Andre Iguodala off the bench has, especially when it is by far the best line-up for the Warriors and he closes the game with 4 starters.

We can check! From 2015-2017, the Warriors at a starter disadvantage with Steph and Iguodala on had a +14.77 net rating, and in those same years at a starter disadvantage with Steph on and Iguodala off, the Warriors had a +11.24 net rating. So, as we’d expect, Iguodala helps, but the numbers were great regardless. And Steph’s player-filtered data here is actually slightly *better* in the 2020s where Iguodala was either not on the Warriors or was an essentially irrelevant player. So obviously Iguodala was a good player, but I don’t think he’s driving the numbers for Steph.

This data is cool, but what exactly does it tell us?

1) We aren't controlling for quality of bench player(s)
2) We aren't controlling for time spent with each bench player(s)
3) We aren't controlling for time spent with starter(s)

These are fair points. But I think I already made relevant caveats regarding these issues in my OP, when I talked about not controlling for the quality of bench players, and also mentioning that this gets at stuff already accounted for in RAPM (which does control for the #2 and #3 things you mentioned). In an ideal world, we might run some sort of starter-disadvantage-specific RAPM such that we are controlling for these things you mention, but I don’t have the time or data to do that.

For example, Mario Chalmers and Chris Bosh are both starters for the 2011 Heat. Nobody needs to ask which 4v5 line-up is better, the one with Chris Bosh as one of the 4 starters or the one with Mario Chalmers.

Furtheremore, the 2007 Spurs [for example], started two centers who played < 20 MPG. Francisco Elson and Fabrico Oberto. Sub in Manu for either one, and boom the Spurs are closing 4v5 in terms of Starters for each team. Is this really impressive? Because the data here is telling us it should be impressive.

I don’t think these are really fair examples of issues with the data I’ve presented, since you raise issues I already explicitly made adjustments for. For the first one, this exact sort of thing is why I presented a version of the data that filtered out someone like Chris Bosh.

That doesn't accurately fix this problem though. A Top 50 player who is a starter and a Top 150 player who is a starter are being treated the same in these scenarios. That's my point.

Take for instance the 2023 Denver Nuggets. Bruce Brown is going to be in almost all their bench units/line-ups as a 29 MPG bench player. He was a low-end starter in terms of quality of player in 2023, yet is a bench piece. He is being treated equal to 2015 Austin Rivers since both count as a player for the "bench", yet neither are providing similar lift/impact as a bench player. That is going to directly and positively effect someone like 2023 Nikola Jokic while negatively and directly effect 2015 Chris Paul.

Do you think the difference in play between Austin Rivers and Bruce Brown is directly related to the level of play 2015 Chris Paul and 2023 Nikola Jokic provide on the court? Does Jokic doing well with Bruce Brown and Chris Paul failing with Austin Rivers [Who wasn't an NBA caliber rotation player in 2015, let alone low-end starter level like Bruce Brown] determine which player [CP3 vs Jokic] is a better "Bench leader"?

We can go deeper. Look at LeBron's numbers in Miami and how they vary year-by-year.

2011: Horrible depth. Non-NBA players left and right playing big-time minutes off the bench. Big Z, Arroyo, Howard and Bibby were notably names playing > 500 Minutes for the team. Hell, Mike Bibby played over 400 Playoff Minutes with a 3.7 PER, 37 TS% and -4.6 BPM .

Now, LeBron's line-ups with these bench players is...bad. Surprising?

2011: -3.3 Net Rtg

But, look what happens when the Heat build out a competent bench, adding in an actual NBA player like Shane Battier to the rotation along with the return of Healthy Udonis Haslem.

2012: +20.6 Net Rtg
2013: +21.4 Net Rtg

Then we get to 2014 and it looks closer to 2011 than 2012 and 2013, with Rashard Lewis and Michael Beasley playing minutes, Wade missing ~30 games and forcing Battier into a starting role [where he really was optimized as a 6th/7th man like in 2012 and 2013 at this point in his career].

2014: +6.9 Net Rtg

It seems like there is far more variance in the actual bench players and their minutes than the actual main player themselves [whether it's Manu, LeBron, CP3 or Jokic]. That's the point I am trying to get at here with these line-ups. The samples are small and the degree at which a bench is talented or less talented is so finicky. Often time one injury, moving your 6th man to the starting role and forcing a team to play a 5-10 MPG player 15-20 MPG is going to have drastic effects on this exercise since the numbers we are working with are smaller-sample.

[lessthanjake]

Very cool data, haven't seen this sort of thing analyzed before so I appreciate you sharing.

This sort of thing is definitely tricky because of all the potential adjustments you can make; the Klay thing, as has been pointed out, and the fact that removing him makes Steph's numbers look far better (and decreases the sample size further to like, half a season's worth of possessions I think?). Lot of different ways to try to adjust things and get very different values. Of course it seems like no matter how you slice it, Manu's numbers are ridiculously impressive.

I calculated Steph's numbers with Klay off the court, but those have already been posted. Also LeBron 2009-13: net rating with a starter disadvantage and sans Wade/Bosh is +10.99.

Yeah, agreed. One can make lots of different filtering decisions and it can swing the numbers (as we saw with the Klay thing). It takes some time to run these, so the only numbers I ran for any of these guys are exactly what I’ve reported out. So I don’t really know what further filtering would do in specific cases, but I suspect it could change things (even if only just due to randomness). What I ran was just the filtering I thought made the most sense, without knowing beforehand how those filtering decisions would affect the data.

And yeah, I actually think that LeBron’s 2009-2013 numbers being really great here in a sense tends to validate that the method is measuring something meaningful, since it does show LeBron’s consensus best years as being the ones where he looks best in this analysis (even down to 2011 being a down year in that timeframe). I take that as an encouraging sign about the analysis.

That said, I’d note again that lumping in 2009 and 2010 with the players-filtered-out version is a bit iffy, since filtering out stars inherently means more of these minutes are with bench or deep bench guys, and not filtering them out is very likely going to leave a lot larger a proportion of the minutes being with the other starters (but just in starter states of 4v5, 3v5, and 3v4). And, without filtering out anyone, that’d include often having the next best guys on the team on (which obviously isn’t the case with years where star teammates are filtered out). So it’s a bit comparing apples and oranges to compare filtered and non-filtered years. But, as I said in my OP, there’s not an obvious person to filter out in those years, so I don’t really have a better solution. My inclination is to think that any filtering in 2009 and 2010 probably wouldn’t change the bottom-line number for 2009-2013 much anyways, since it’d probably make the 2009 and 2010 numbers less good but also make those years a smaller part of the sample, leaving the super high 2012 and 2013 numbers as a larger part of the data. For instance, if we filtered out Mo Williams and Varejao, we get +24 in 316 possessions in 2009, and +23 in 233 possessions in 2010. Those unsurprisingly look less good than the non-filtered numbers for those years do, but the overall 2009-2013 time period ends up being +10.34 anyways, which is barely lower than what you listed above and is still outrageous.

I will note that Manu has timeframes that are pretty outrageous too. For instance, his 2012-2016 span with stars filtered is a +13.75 net rating.

There's no way to do this cleanly once you get into those adjustments, which is kinda what this shows. Yeah, if you don't filter out anyone for the first stint Cavaliers then you're not filtering out LeBron's best teammates.

But comparing the that first stint Cavs lineups to the disadvantaged Steph lineups shows that you're obviously not leaving them with the same level of non-starting teammates

- The filtered LeBron first stint Cavs lineup with the most possessions in a disadvantaged starter state is LeBron, Gibson, Hickson (20-years-old), Szczerbiak (out of league next year), Pavlovic. +8 in 30 possessions

- Next up is LeBron, Gibson, Hickson, Kinsey, Szczerbiak. Don't even remember hearing of Tarence Kinsey lol, apparently he was also out of the league after the 2009 season (tore it up in Europe though). +9 in 26 possessions

While filtered Steph lineups could be something like:

- Steph/Poole/Klay/Wiggins/Looney (+10 in 43 possessions)
- Steph/CP3/Klay/Kuminga/Looney (+19 in 20 possessions)
- Steph/GP2/Bjelica/OPJ/Iguodala (+11 in 46 possessions)

A problem with RAPM is interpretability. That's why being able to say "the Spurs had a +6.00 net rating from 2003-2017, at a starter disadvantage and Ginobili on and Duncan, Parker, Robinson, and 2014-onwards Kawhi all off" is pretty cool, and certainly an interesting tidbit. But there are concerns with a comparative approach, and I think LeBron/Curry is an extreme example that illustrates that (first stint Cavs dearth of talent vs Strength in Numbers). Like you said, comparing filtered to non filtered years is apples to oranges, but so is the bench players of one team compared to the bench players of another. That's where RAPM comes in and helps. Not to say you did anything wrong, just something to consider.

Yeah, I don’t disagree with any of this, which is why I caveated upfront in my OP that “none of this analysis corrects for the precise quality of the role players or bench guys on each team.” As I mentioned in a post just above, in an ideal world, we might run some sort of starter-disadvantage-specific RAPM that can control for this sort of stuff, but it’s not something that’s possible. So what we are left with is something that has the same sorts of flaws that on-off has in general, except probably a bit less so once we’re filtering out great players and talking about starter-disadvantage—since that leaves us with role/bench players and there’s generally not going to be so big a difference in the quality of those sorts of players, though obviously Ginobili himself is a very prominent exception where a bench player was incredibly good! But the fact that there is inherently at least some difference (since players don’t have the same role players on their teams) is, of course, a limitation/flaw in the analysis. That’s part of why I focused first and foremost on a comparison with Duncan, since that does actually materially limit how much this issue matters (though even that’s not perfect, because we can’t assume they had the exact same distribution of what role players they played these minutes with).

[lessthanjake]

And it’s the type of thing that could be nitpicked about most teams. For instance, LeBron’s data on the Heat is surely sweeping in minutes where Joel Anthony or Udonis Haslem (who were starters) are off and Shane Battier (a better player, who wasn’t a starter) was on. And the same sort of thing could be true of opponents, who might have bench players that are better than their starters, leading to their minutes with a starter advantage not actually being as strong as we’d think. It’s not wrong to point out that starter state isn’t a perfect proxy for the relative quality of the players on the court. The fact that it’s not a perfect proxy is inherently a blind spot in the data I’ve presented. But I don’t think this is at all specific to the Spurs.

It isn't specific to the Spurs, THAT IS MY POINT. My point is the samples are small, the difference between the quality of players like "Shouldn't be playing any minutes" and "Good 6th man" is massively large. None of your data accounts for the massive differences in quality of benches.

You can't just say "Well everyone has asterisks so let's just ignore them" because that isn't a logical way to use this data.

You’re just repeating things that I explicitly said upfront in my OP. Read the first two caveats listed at the bottom of my OP. You’re just saying the same things. I fully understand the limitations of the data, and I clearly don’t disagree with you since I said what you’re saying before you did. I still think it is interesting information. We can essentially look at it with the same sorts of caveats that we look at something like single-season or multi-season on-off data with. It’s interesting, but cannot capture all the context.

[lessthanjake]

NOTABLE EDIT

I realize I made an error with LeBron’s data here, specifically for the player-filtered years he was in Miami. I cannot even figure out what I did wrong (which makes me a little antsy), but I reran it to check something Colbinii had posted and discovered that the data I reported in those Miami years is wrong. Whatever issue caused this does not seem to have affected any of the other player-filtered data for LeBron, since I checked the second-stint Cleveland and Lakers years again and everything there checked out. Here is the corrected data for LeBron in Miami:

LeBron James with a Starter Disadvantage - Without Wade or Bosh

2011: +43 in 573 possessions
2012: -6 in 548 possessions
2013: +64 in 630 possessions
2014: -20 in 635 possessions

I will fix this in the OP and in the raw data post. Overall, this corrected data is not as good for LeBron as the erroneous data had been.

In general, I’d be happy for others to check my work. It’s certainly quite possible this wasn’t the only error I made, and I’m a little nervous about the fact that I can’t even figure out what the mistake in the initial data even was. Running both player-filtered and not player-filtered numbers for so many seasons for a bunch of players, and on top of that having to do some additional filtering and changes to the “Starter State” parameters to make Ginobili always be treated as a starter must’ve resulted in me making some sort of mistake along the way, especially in light of the fact that PBPstats can be glitchy sometimes. Hopefully it was just a one-off thing.

[lessthanjake]

I wonder what advantage someone playing with, say, 2015-2017 Andre Iguodala off the bench has, especially when it is by far the best line-up for the Warriors and he closes the game with 4 starters.

We can check! From 2015-2017, the Warriors at a starter disadvantage with Steph and Iguodala on had a +14.77 net rating, and in those same years at a starter disadvantage with Steph on and Iguodala off, the Warriors had a +11.24 net rating. So, as we’d expect, Iguodala helps, but the numbers were great regardless. And Steph’s player-filtered data here is actually slightly *better* in the 2020s where Iguodala was either not on the Warriors or was an essentially irrelevant player. So obviously Iguodala was a good player, but I don’t think he’s driving the numbers for Steph.

This data is cool, but what exactly does it tell us?

1) We aren't controlling for quality of bench player(s)
2) We aren't controlling for time spent with each bench player(s)
3) We aren't controlling for time spent with starter(s)

These are fair points. But I think I already made relevant caveats regarding these issues in my OP, when I talked about not controlling for the quality of bench players, and also mentioning that this gets at stuff already accounted for in RAPM (which does control for the #2 and #3 things you mentioned). In an ideal world, we might run some sort of starter-disadvantage-specific RAPM such that we are controlling for these things you mention, but I don’t have the time or data to do that.

For example, Mario Chalmers and Chris Bosh are both starters for the 2011 Heat. Nobody needs to ask which 4v5 line-up is better, the one with Chris Bosh as one of the 4 starters or the one with Mario Chalmers.

Furtheremore, the 2007 Spurs [for example], started two centers who played < 20 MPG. Francisco Elson and Fabrico Oberto. Sub in Manu for either one, and boom the Spurs are closing 4v5 in terms of Starters for each team. Is this really impressive? Because the data here is telling us it should be impressive.

I don’t think these are really fair examples of issues with the data I’ve presented, since you raise issues I already explicitly made adjustments for. For the first one, this exact sort of thing is why I presented a version of the data that filtered out someone like Chris Bosh.

That doesn't accurately fix this problem though. A Top 50 player who is a starter and a Top 150 player who is a starter are being treated the same in these scenarios. That's my point.

Take for instance the 2023 Denver Nuggets. Bruce Brown is going to be in almost all their bench units/line-ups as a 29 MPG bench player. He was a low-end starter in terms of quality of player in 2023, yet is a bench piece. He is being treated equal to 2015 Austin Rivers since both count as a player for the "bench", yet neither are providing similar lift/impact as a bench player. That is going to directly and positively effect someone like 2023 Nikola Jokic while negatively and directly effect 2015 Chris Paul.

Do you think the difference in play between Austin Rivers and Bruce Brown is directly related to the level of play 2015 Chris Paul and 2023 Nikola Jokic provide on the court? Does Jokic doing well with Bruce Brown and Chris Paul failing with Austin Rivers [Who wasn't an NBA caliber rotation player in 2015, let alone low-end starter level like Bruce Brown] determine which player [CP3 vs Jokic] is a better "Bench leader"?

We can go deeper. Look at LeBron's numbers in Miami and how they vary year-by-year.

2011: Horrible depth. Non-NBA players left and right playing big-time minutes off the bench. Big Z, Arroyo, Howard and Bibby were notably names playing > 500 Minutes for the team. Hell, Mike Bibby played over 400 Playoff Minutes with a 3.7 PER, 37 TS% and -4.6 BPM .

Now, LeBron's line-ups with these bench players is...bad. Surprising?

2011: -3.3 Net Rtg

But, look what happens when the Heat build out a competent bench, adding in an actual NBA player like Shane Battier to the rotation along with the return of Healthy Udonis Haslem.

2012: +20.6 Net Rtg
2013: +21.4 Net Rtg

Then we get to 2014 and it looks closer to 2011 than 2012 and 2013, with Rashard Lewis and Michael Beasley playing minutes, Wade missing ~30 games and forcing Battier into a starting role [where he really was optimized as a 6th/7th man like in 2012 and 2013 at this point in his career].

2014: +6.9 Net Rtg

It seems like there is far more variance in the actual bench players and their minutes than the actual main player themselves [whether it's Manu, LeBron, CP3 or Jokic]. That's the point I am trying to get at here with these line-ups. The samples are small and the degree at which a bench is talented or less talented is so finicky. Often time one injury, moving your 6th man to the starting role and forcing a team to play a 5-10 MPG player 15-20 MPG is going to have drastic effects on this exercise since the numbers we are working with are smaller-sample.

Please note that, as I mentioned in my above post, I realized that I’d made some sort of error in LeBron’s player-filtered data for Miami. I have no idea what the error was. But it’s worth noting that the corrected data doesn’t actually match the rhetorical point you’re making here, with LeBron doing significantly better in starter-disadvantage minutes without Wade or Bosh in 2011 than he did in 2012. Of course, on a year by year basis these aren’t large samples, so that’s not all that surprising even if we think the 2012 role players were better. But it seemed worth noting (as is the general fact that a portion of the data had been wrong).

[lessthanjake]

Another somewhat related piece of data I want to add here:

One of the bits of skepticism I’ve seen regarding Ginobili is that he was just farming the opposing teams’ bench players. To some extent, as a bench player he likely *was* facing opponents’ bench players a larger proportion of the time than starters would. But, of course, he was also playing with bench players more. And the starter-disadvantage data addresses this to some degree, since it’s clear that he did really well even when his team was actually the one with more bench guys in the game.

But another interesting way to look at this issue is to filter the “Starter State” data down to only situations in which Ginobili was facing all five starters on the other team. This gets at how the Spurs did with Ginobili when they weren’t facing opponents’ bench players at all.

What do we find there?

Over the course of Ginobili’s entire career (2003-2018), when facing all 5 starters on the other team, the Spurs had a net rating of +11.37 with Ginobili on the court.

I haven’t checked how other players fare in this, but I’m inclined to think that that is a remarkable number. And what’s interesting is that PBPstats has his career ON value at +10.01. So Ginobili’s net rating was actually *better* against the opponents’ starters than it was overall! Obviously, this is surely caused by the fact that the lineups he was playing with in those situations likely tended to be stronger than average too. In the sense that there’s probably a ton of Duncan in these minutes, it’s not super surprising that it’d be higher than Ginobili’s average. But it does demonstrate that Ginobili wasn’t just farming high ON values from playing against bench players. His lineups destroyed the opponents’ starters too!

I also ran this data for the playoffs, and in the playoffs, Ginobili’s lineups were +5.66 when facing all 5 starters on the other team. That is actually a bit below his career playoff ON rating of +7.49, but it’s still really high. Considering how good the starting five of the typical playoff opponent is (especially in the Western Conference in the era Ginobili played in), that’s a super high number!

__________________

EDIT: For reference, as a flip side to this analysis, I ran numbers for how the Spurs did with Ginobili on and the opponent having 3 or fewer starters on the court—to basically get at what Ginobili’s teams did when facing multiple bench players. The result was that, for his career, Ginobili had a +10.31 net rating against teams with 3 or fewer starters on the court. His career ON value is +10.01, so this is slightly above that, but not in any meaningful way. So, as a factual matter, I’d say it’s not the case that Ginobili’s net rating numbers are being driven by facing bench lineups. He actually had a higher net rating against all 5 starters on the other team than against 3 or fewer starters! Even if we filter it down to when the opponent only had 2 or fewer starters on the floor, Ginobili’s net rating is +11.65. Of course, that’s higher than his overall net rating, but not by a huge amount, and it’s actually virtually identical to his net rating against all 5 starters on the other team. Obviously, it is surely the case that these numbers are all so similar because the Spurs lineups with Ginobili in these scenarios were different (i.e. they surely tended had better Spurs on the court with Ginobili against more starter-heavy opposing lineups), but I think we can pretty clearly say that Ginobili’s super high career net rating isn’t simply a product of facing bench lineups.

[lessthanjake]

I ran these starter-disadvantage numbers for Shaq and Kobe. The fact that the data starts at 2001 is a bit unfortunate, but we get a fair few years for both. For Shaq, I stopped at 2007, and for Kobe I stopped at 2013.

Here’s the data:

Shaquille O’Neal Plus Minus at a Starter Disadvantage

2001: +33 in 1113 possessions
2002: +226 in 1336 possessions
2003: +113 in 1003 possessions
2004: +102 in 1083 possessions
2005: +22 in 798 possessions
2006: -31 in 673 possessions
2007: -31 in 433 possessions

For 2001-2007, this is a +6.74 net rating. If we just look at 2001-2005 (which is arguably the only stuff that is still prime Shaq), it is a +9.30 net rating at a starter disadvantage.

Kobe Bryant Plus Minus at a Starter Disadvantage

2001: -22 in 1253 possessions
2002: +151 in 1478 possessions
2003: +81 in 1453 possessions
2004: +19 in 1069 possessions
2005: -96 in 1015 possessions
2006: +26 in 1254 possessions
2007: +63 in 1591 possessions
2008: +96 in 1341 possessions
2009: +155 in 1306 possessions
2010: +20 in 840 possessions
2011: -52 in 764 possessions
2012: +90 in 910 possessions
2013: +5 in 1230 possessions

For 2001-2013, this is a +3.46 net rating at a starter disadvantage. There’s a peak five-year period there from 2006-2010 that gets up to a +5.69 net rating.

What if we filter out star teammates? For Shaq, it’d make sense to filter out Kobe and Wade. For Kobe, I’ll filter out Shaq and Pau Gasol. For Kobe, this does leave a few years in the middle where no one is being filtered out, which is comparing apples to oranges to some degree, but like with LeBron’s first Cleveland stint there’s no obvious person to filter out.

Shaquille O’Neal Plus Minus at a Starter Disadvantage - Without Kobe or Wade

2001: +24 in 422 possessions
2002: +85 in 572 possessions
2003: -33 in 250 possessions
2004: +7 in 474 possessions
2005: +20 in 319 possessions
2006: -31 in 242 possessions
2007: -18 in 346 possessions

Overall, for the whole 2001-2007 time period, that is a +2.02 net rating for Shaq at a starter disadvantage without Kobe or Wade. If we limit it to just 2001-2005, it is up to a +5.06 net rating. Not a big sample on any of this though.

Kobe Bryant Plus Minus at a Starter Disadvantage - Without Shaq or Gasol

2001: -37 in 562 possessions
2002: +10 in 714 possessions
2003: -65 in 700 possessions
2004: -76 in 460 possessions
2005: -96 in 1015 possessions
2006: +26 in 1254 possessions
2007: +63 in 1591 possessions
2008: +67 in 1175 possessions
2009: -22 in 360 possessions
2010: +45 in 318 possessions
2011: -63 in 231 possessions
2012: -8 in 224 possessions
2013: -14 in 778 possessions

Kobe’s total net rating at a starter disadvantage without Shaq or Gasol from 2001-2013 was -1.81. Pretty disappointing stuff. There is a five-year span there from 2006-2010 where his net rating in these situations was +3.81, and that probably does generally correlate with his actual peak as a player.

I also ran the numbers for Shaq and Kobe together at a starter disadvantage:

Shaq and Kobe together at a Starter Disadvantange

2001: +15 in 691 possessions
2002: +141 in 764 possessions
2003: +146 in 753 possessions
2004: +95 in 609 possessions

Overall, that’s comes out to a +14.09 net rating with Shaq and Kobe on at a starter disadvantange. Which is actually even above what we saw from 2014-2024 Steph and Draymond together! That said, since this is only four years, the sample size on this isn’t huge, and Steph + Draymond and Duncan + Ginobili did have higher net ratings at a starter disadvantage in similar spans.

[Aleco]

Can u do minutes instead of possesions