Retro Player of the Year 1992-93 UPDATE — Hakeem Olajuwon

[lessthanjake]

A season's worth of games is nowhere near enough to stabilize and as has been noted, applying a similar sampling distributions and process for other players with full rapm out led to big jumps. When we have similar data for everyone, this becomes useful. Right now it's not.

I think you may be conflating concepts here a bit. A season’s worth of games is definitely not enough to “stabilize” RAPM for purposes of minimizing random variance. But we aren’t talking about random variance. We’re talking about how long it takes for it to stabilize in terms of moving a player’s RAPM from the 0 prior to the general vicinity of where they belong. That’s not about random variance at all—quite the opposite actually. It’s a fundamentally different concept.

And a single season’s worth of data is a pretty good amount for these purposes. If it wasn’t then we’d expect the top pure RAPM numbers in a given year to be consistently higher as we expand out the number of years being analyzed. And we don’t really see that. Take, for instance this website, which is pure RAPM (without any priors, just like Squared’s RAPM) that conveniently allows us to filter the top players by one-year, three-year, and five-year RAPM: https://thebasketballdatabase.com/2016-17RegularSeasonPlayerRAPMComprehensive.html. Obviously the exact details will depend on the year we look at and there’s random variance at play here, but if you peruse the years you’ll find that the top values for one-year RAPM are generally very similar to the top values for three-year and five-year RAPM.

Of the top 10 in your site, 7 see at least a >1 point drop from 1-year to 5 year and 3 see a >2 point drop. (2 see <.5 increases). These results look pretty different from 1 year to 5 year and those are full samples where it's available for everyone.

This is a really curious post, since your argument is that RAPM goes substantially *up* as the sample gets larger than one year, and here you’re pointing to the fact that a bunch of peoples’ RAPM values went *down* as the sample got larger than one year. I think you got yourself confused here, and argued against your own point.

I’ll note that we see similar things in others years too. For instance, to take an adjacent year: https://thebasketballdatabase.com/2015-16RegularSeasonPlayerRAPMComprehensive.html. Even if you filter it for the top five-year RAPMs (so we’re not ending up accidentally cherry-picking positive random variance in the one-year data), 7 of the top 10 had higher one-year RAPMs. That said, having briefly perused other years before I made my prior post, I think there’s years where it goes the other way (random variance is at play here, so we’d expect that), which is why I conservatively said the values “are generally very similar” rather than making any actual directional assertion. Ultimately, the point is merely that we do not see the longer timespans be consistently far higher—and we’d need to see them be consistently higher if one year was nowhere near large enough for pure RAPM to stabilize for these purposes (and Hakeem needs that to be the case given how much higher Jordan’s Squared RAPM is). You arguing that these RAPM values get notably *lower* as the sample gets larger definitely strongly suggests RAPM in longer timespans are not consistently far higher than one-year RAPM!

[Djoker]

I love how apparently an entire season worth of games isn't nearly enough to stabilize RAPM but then at the same time WOWY off samples of anywhere between 5 games and 24 games are used as end all be all evidence that Lebron provides unsurpassed lift.

[AEnigma]

I must have missed where the stance was ever expressed that there is a linear relationship in any direction between a sample size and its output number. The relationship at play is the smaller or less complete the sample, the less likely it is to be representational. Which has been echoed over and over and over and over again by dozens of different people, to no avail.

I love how apparently an entire season worth of games isn't nearly enough to stabilize RAPM but then at the same time WOWY off samples of anywhere between 5 games and 24 games are used as end all be all evidence that Lebron provides unsurpassed lift.

And then this is not only off-topic, but an explicitly bad faith misrepresentation.

[Special_Puppy]

I think the issue is we are trying to solve many of these comparisons with data that is very incomplete. That's not going to be a fool proof or entirely convincing way to do it. None of us are new to these comparisons or the data that can support them for the most part. I think this would work better if people made their own points and there wasn't an appeal to just incomplete data or box score composites as a form of rebuttal. You can't use something like box score composites to compare a center who has a case for being the most impactful defender after Russell to a guard who scores a lot and likely got way too many dpoy votes. It created the illusion that they are similar defensively imo which is part of the problem in comparing these guys 30 years later. There's going to be some gray area that each voter or poster has to fill in in whichever way they like in these discussions. More so the less reliable data we have(which is never going to check every box that people vote according to anyhow). Same as how people will weigh a guy winning a ring in a given year differently also.

What are we left with without box score composites and stuff like RAPM over the larger time period? Just WOWY, film studies, and slashlines?

[lessthanjake]

I must have missed where the stance was ever expressed that there was a linear relationship in any direction between a sample size and its output number. The relationship at play is the smaller the sample, the less likely it is to be representational. Which has been echoed over and over and over and over again by dozens of different people, to no avail.

I love how apparently an entire season worth of games isn't nearly enough to stabilize RAPM but then at the same time WOWY off samples of anywhere between 5 games and 24 games are used as end all be all evidence that Lebron provides unsurpassed lift.

And then this is just an explicitly bad faith misrepresentation.

I think this *has* actually been asserted in this thread, and has also been asserted in the past. And it’s actually a valid point, depending on what type of RAPM model we’re looking at. RAPM models start everyone at a prior, and it takes some data for the model to get confidence to move someone far away from that prior. For most RAPM models, this doesn’t lead to a predictable effect in one direction as the sample goes up. This is because most RAPM models involve some sort of box prior. So, whether a player’s RAPM goes up or down as the sample goes up depends on whether the box prior underestimated or overestimated the player’s true value. However, Squared RAPM is pure RAPM, with no prior. In other words, everyone starts at zero. Since we can generally assume that the true RAPM value for star players is above zero, this means that with no prior, we can expect that a star player’s RAPM will initially go up as the sample goes up. I don’t think the question is whether this is true for Squared RAPM. It almost certainly is. The question for these purposes is just how large the sample has to be before this effect loses its importance (i.e. how much data is needed before star players will tend to have reached the general vicinity of their true value and stop inevitably going up as the sample increases). That point comes eventually. Regardless of a prior of zero, players’ RAPM won’t just infinitely go up as the sample increases. My point is that having roughly a single season’s of data is a pretty good amount for these purposes, as evidenced by the fact that we can look at no-prior RAPM values over different timespans and see that the top one-year values tend to generally on average be very similar to the top three-year and five-year values. OhayoKD responded to me pointing this out, by noting that many of the top five-year values are lower than the top one-year values, which obviously tends to suggest one-year is enough data for these purposes rather than the opposite.

And, to be clear, the import of this here is that in Squared’s full-sample RAPM, Hakeem has a RAPM that is only a bit more than half of Jordan’s. The response to that from OhayoKD was basically that this is because Jordan’s sample is larger. But Hakeem’s sample is roughly an entire season’s worth of data, so if that’s a pretty good amount for these purposes, then it would not seem all that plausible for that to be the reason Hakeem’s RAPM was so much lower than Jordan’s.

All that said, while the above is relevant given that discussion on this thread has centered on Jordan and Hakeem, obviously Squared’s full-sample RAPM data is from more than just 1993. Squared does have 1993-specific data in which Jordan is far ahead of Hakeem, but I’m inclined to think that it is a valid point that for just 1993 specifically the Hakeem sample is small enough that we’d think this effect would probably be at play.

[AEnigma]

Generating complete RAPM for a single season (let alone across multiple) requires hundreds of games, not most of a season for one player and a scattering of them for others.

[penbeast0]

Oh ok so you're waving everything away. Your wings... um hands... must be getting tired.

We have full data ON-OFF from Pollack so we have RAPM for the 1994-1996 period and Hakeem doesn't look that good either.

1994: 4.03 RAPM - 4th behind Robinson, Willis, Malone
1995: 3.44 RAPM - 6th behind Robinson, Shaq, Malone, Hardaway, Pippen
1996: 3.04 RAPM - 11th behind Jordan, Robinson, Hardaway, Pippen, Malone, Stockton, Shaq, Kukoc, Hill, Grant

Nothing that shows some force majeure type of impact here. Jordan on the other hand has the highest career playoff ON-OFF with a full sample, one of if not the highest regular season ON-OFF with a partial sample and leads in partial sample RAPM in 1988, 1991, 1993, 1996 as well in full sample RAPM in 1997.

Hakeem is a better player than Robinson and Malone because those two guys shrink in the PS while Hakeem rises in the PS but compared to Jordan, there just isn't a very serious case to be made whether you use box score or impact data.

One thing I’d note is that, despite having full on-off data, I don’t think we really have RAPM for 1994-1996, because we don’t actually know all the lineup iterations at all times like we would need to in order to have RAPM. Am I missing something? I seem to vaguely recall that there is something floating out there for those years that styles itself as “RAPM” but isn’t actually RAPM.

Regarding the point about Squared’s partial samples, I do think there’s a valid point that players that happen to have larger samples can be advantaged, because the prior in Squared’s analysis is 0, so if the sample is too small then it basically won’t have enough data to have moved as far away from the 0 prior as it should. So, as the sample for a great player stops being small, we will tend to see the RAPM increase. This isn’t some infinite effect, though. Obviously, a player’s RAPM doesn’t just go up constantly as you increase the sample. This stabilizes when it has a decent-sized sample (though obviously random noise is still at play), which is probably why multi-year RAPM values aren’t just inevitably higher than single-year RAPM values. So this isn’t some always-applicable excuse. And that’s an important point here. As I once noted on this topic particularly regarding Hakeem, when this argument was brought up about the multi-year RAPM that Squared put together:

Spoiler:

We have like a full season’s worth of games for Hakeem in Squared’s sample. Again, the point you’ve noted about larger sample sizes helping players is certainly a relevant thing when there’s no prior because it will take time for a great player’s RAPM to get up to where it should be, but star players’ RAPM doesn’t just keep going up infinitely as you add more games. Squared’s stuff is incomplete and not conclusive, but I think we can make a very good inference that it’s not just a sample-size-with-no-prior thing when we have like a full season’s worth of games for Hakeem and his RAPM is still only a bit above half Jordan’s. Furthermore, we also have Engelmann’s quarter-by-quarter RAPM approximation for the 1990s, which has Jordan *way* above Hakeem. We also have actual RAPM from 1997 and 1998, which has Jordan way above Hakeem. Granted, I think we could validly define 1998 as not a prime year for Hakeem, but 1997 definitely still was and their RAPM wasn’t close. We also have WOWYR putting Jordan way above Hakeem, as well as Jordan doing better in Moonbeam’s related analysis. The bottom line is that, while we certainly don’t have perfect impact data for them, we have enough to say that the impact data seems to validate the public perception from the time period that it wasn’t particularly close.

Spoiler:

The bottom line is that if Hakeem were similar to Jordan in RAPM, we’d expect him to be substantially closer in the Squared data, given the size of the sample we have for Hakeem. It’s not some tiny sample where there’s not enough games for the model to at all figure out how impactful he is. Just for reference, the individual season RAPM that Squared has done for Jordan includes lots of seasons where the sample for Jordan in terms of possessions is similar or smaller than the overall sample he has for Hakeem. And, aside from in Jordan’s rookie season with a sample size less than half of Hakeem’s overall sample, Jordan’s single-season RAPM is still always higher than what Hakeem has in Hakeem’s overall sample (and Jordan’s very close even in that rookie year). For instance, Jordan’s 1988 RAPM is up at +7.47, despite having just 3595 possessions in the sample, compared to Hakeem’s +5.43 in 6066 possessions in the overall sample. Granted, it’s a bit problematic to compare precise numbers spit out by two different regressions, but it’s pretty clearly possible in Squared’s RAPM calculations to have notably higher RAPM than Hakeem got overall, in similar or smaller samples than he has for Hakeem. Which strongly suggests that Jordan being far ahead isn’t just a result of Hakeem having a smaller sample. Obviously Squared’s stuff always comes with the caveat that it’s partial data so isn’t conclusive, but we can certainly draw a fairly good probabilistic inference from what we have that Hakeem wasn’t nearly as impactful as Jordan. And that’s especially true when we look at it in context with all the other data we have that says similar things, discussion of which you conspicuously cut from your quote of me under the guise of making an immature snarky comment. And this data backs up what should be our baseline prior that Jordan was substantially better, based on contemporaneous perception of the two at the time.

Relatedly, we have the fact that Squared’s individual-season data has similarly large samples for Magic as for Jordan in seasons where their primes overlapped. Obviously, this isn’t the case for 1993 because Magic didn’t play, but it’s worth noting that Jordan’s Squared RAPM was higher than Magic’s in 1988, despite the fact that the sample for Magic was larger. And that’s in a year where this renewed POY vote gave Magic the POY. Meanwhile, in Squared’s 1991 RAPM, Jordan has a larger sample than Magic but not by much, and Jordan’s RAPM is way higher (60% higher). And that’s in a year where the poster you’re responding to voted Magic over Jordan. I mention this stuff about Magic because I think we can all recognize that Magic Johnson is almost certainly an impact monster, and yet Squared’s RAPM had prime Jordan above prime Magic, with this sample-size argument clearly not being the reason.

Apologize if this has been linked already, but the entire Squared2020 RAPM sample also has Jordan comfortable in number 1. Jordan has the largest sample though so there’s the caveat he’s getting regularized the least

Read on Twitter

?s=46&t=kXz47zZ_dZNjyZ0pjPEBWQ

Why isn't Karl Malone on Squared2020's RAPM sample. Stockton and Eaton are. Is he really that low or is it an omission?

[lessthanjake]

Generating complete RAPM for a single season (let alone across multiple) requires hundreds of games, not most of a season for one player and a scattering of them for others.

Okay, so you’d lobby to entirely reject usage of Squared’s RAPM because it’s not based on fully complete data. You’re free to argue that, of course. That’s not the argument OhayoKD was making, though.

If you want to entirely reject Squared’s RAPM (rather than taking it into account while acknowledging it’s not as reliable as full data would be—which I think is the more common position here on it), then that raises the question of what information you don’t reject. We don’t have “complete RAPM” for this era. Beyond Squared’s RAPM, we also have J.E.’s RAPM approximation using quarter-by-quarter results. That data is complete, but the methodology is necessarily more imprecise. It similarly finds Jordan far ahead of Hakeem (Jordan’s gets over twice Hakeem in this measure). We also have complete RAPM for 1997 (and 1998 FWIW), which also has Jordan far ahead of Hakeem. Same with WOWYR, as well as Moonbeam’s very similar analysis. And box data—including those that have been designed to correlate with impact—also have Jordan ahead of Hakeem. There’s just a lot of stuff that needs to be rejected here. It’s not just Squared’s RAPM.

[Cavsfansince84]

What are we left with without box score composites and stuff like RAPM over the larger time period? Just WOWY, film studies, and slashlines?

I'm not saying to completely disregard them. What I am saying is that you have to use some degree of context with them. All metrics require some degree of context to get meaning out of them. They all have blindspots so that's part of using them. Along with other things that you as an individual factor into your opinions. I think the idea is just to do the best you can and trying to be consistent/fair in the approach/methodology you use.

[OhayoKD]

I think you may be conflating concepts here a bit. A season’s worth of games is definitely not enough to “stabilize” RAPM for purposes of minimizing random variance. But we aren’t talking about random variance. We’re talking about how long it takes for it to stabilize in terms of moving a player’s RAPM from the 0 prior to the general vicinity of where they belong. That’s not about random variance at all—quite the opposite actually. It’s a fundamentally different concept.

And a single season’s worth of data is a pretty good amount for these purposes. If it wasn’t then we’d expect the top pure RAPM numbers in a given year to be consistently higher as we expand out the number of years being analyzed. And we don’t really see that. Take, for instance this website, which is pure RAPM (without any priors, just like Squared’s RAPM) that conveniently allows us to filter the top players by one-year, three-year, and five-year RAPM: https://thebasketballdatabase.com/2016-17RegularSeasonPlayerRAPMComprehensive.html. Obviously the exact details will depend on the year we look at and there’s random variance at play here, but if you peruse the years you’ll find that the top values for one-year RAPM are generally very similar to the top values for three-year and five-year RAPM.

Of the top 10 in your site, 7 see at least a >1 point drop from 1-year to 5 year and 3 see a >2 point drop. (2 see <.5 increases). These results look pretty different from 1 year to 5 year and those are full samples where it's available for everyone.

This is a really curious post, since your argument is that RAPM goes substantially *up* as the sample gets larger than one year, and here you’re pointing to the fact that a bunch of peoples’ RAPM values went *down* as the sample got larger than one year. I think you got yourself confused here, and argued against your own point.

No? The argument was about what happens when a player's proportion of a total sample goes up. The more a player dominates a sample, a more they are pulled away from the mean.

The more years you play, generally your portion of the total sample shrinks. To whatever degree there is utility in using representatively distributed samples to model unrepresentatively distributed ones, what you would be looking for is drops.

I love how apparently an entire season worth of games isn't nearly enough to stabilize RAPM but then at the same time WOWY off samples of anywhere between 5 games and 24 games are used as end all be all evidence that Lebron provides unsurpassed lift.

Not really sure where to start here.

1. Lebron's WOWY is not the end-all be all. He is also an era-outlier in RAPM despite spending several prime-years staggering with a similar teammate playing a position up, and also the most tested player situationally, and also an outlier in older and younger years, and also a Hakeem-level postseason riser. Jordan has a shot at one of those things. At the moment, Magic has most of them.

2. A season's worth of RAPM is not 82 games, it is a sample of what will overwhelmingly be 5-8 minute snippets of games stitched together with a few full games here and there. And whatever output comes from there is artificially curved towards 0 as an artificial scale is created to approximate adjustments to said 5-8 minute snippets. And in this case it is a sample dominated by one team that like most good teams of the era platooned a bunch. And in this season you are using it against a player who has 1/6th of the sample.

The main benefit of RAPM is it's stabler, but the trade-off of stability is you are no longer tracking real season to season impact, you are tracking a baseline. Even if you want to take that baseline at face-value, the 1-year outputs are artificially fitted to that baseline.

That actual stretches of 82 or near 82 games consistently have Jordan's teammates much better than the bits based on him missing a few minutes a game is a pretty big hint lineup effects are at play here.

[lessthanjake]

Of the top 10 in your site, 7 see at least a >1 point drop from 1-year to 5 year and 3 see a >2 point drop. (2 see <.5 increases). These results look pretty different from 1 year to 5 year and those are full samples where it's available for everyone.

This is a really curious post, since your argument is that RAPM goes substantially *up* as the sample gets larger than one year, and here you’re pointing to the fact that a bunch of peoples’ RAPM values went *down* as the sample got larger than one year. I think you got yourself confused here, and argued against your own point.

No? The argument was about what happens when a player's proportion of a total sample goes up. The more a player dominates a sample, a more they are pulled away from the mean.

The more years you play, generally your portion of the total sample shrinks. To whatever degree there is utility in using representatively distributed samples to model unrepresentatively distributed ones, what you would be looking for is drops.

I don’t think this is the point Bad Gatorade was making in what you quoted, nor does it make much of any sense to me (while the related point I take Bad Gatorade as making there that I addressed does actually make good sense, for reasons I’ve explained in this thread). I think you’re either misunderstanding a point someone else made, or just not wanting to admit you goofed and are hand-waving to avoid doing so. And, in any event, in the pure RAPM I provided that you were commenting on, there’s not even any reason to think players’ proportion of the total sample would necessarily go up or down as the number of years goes up. Your assertion to the contrary makes no sense. In reality, that would just depend on how much time the player happened to miss in the different seasons. If a player missed some time in the one season in the one-year sample, but missed very little time in the seasons further back, then they’d make up a lower proportion of the total sample in the one-year sample, and vice versa if you flipped that. So your observations about the values in the RAPM source I provided really don’t map onto this argument you now claim to be making, even if that argument theoretically made sense (which I don’t see how it does and you’ve certainly not explained in any meaningful way). This new argument you’re purporting to make would be something that the RAPM source I provided wouldn’t really tell us much about either way, rather than telling us something about it using the information from it that you pointed out (which is part of why I think you’re just hand-waving to avoid admitting you goofed—the point you’d tried to make doesn’t really appear to logically map onto what you now claim to have been arguing).

[AEnigma]

No, you keep trying to attribute nonsense arguments to him based on slight imprecisions in his choice of language — e.g. offhandedly saying “big jumps” must mean his stance is that by rule RAPM increases with increased sampling, and thus if he points out where several players saw declines with increased sampling, then he is now arguing against himself. Either he did not notice you doing that, or he is charitably letting it be.

There is no sincere reason to spend pages pushing incomplete “RAPM” with a predominant skew to one player/team.

[B-Mitch 30]

I plan to talk about this more in my voting post (assuming the book I need to round it out arrives in time), but I thought the Hornets had one of the best offenses this year. Only playoff team besides the Jazz, Pacers, and Spurs to be in the top 10 of eFG, turnover percentage, and free throws per field goal attempt. Of those teams, only the Pacers were better at offensive rebounding. Larry Johnson is clearly the best player in Charlotte, leading the team in minutes, about tying Alonzo Mourning in offensive rebounds, having some of the most assists as a power forward in the NBA, and scoring 22 points per game on great efficiency. Him and the team were also decent in the playoffs, though obviously no match for the Knicks. What do you think about his and the team's season?

[One_and_Done]

I plan to talk about this more in my voting post (assuming the book I need to round it out arrives in time), but I thought the Hornets had one of the best offenses this year. Only playoff team besides the Jazz, Pacers, Spurs, and Suns to be in the top 10 of eFG, turnover percentage, and free throws per field goal attempt. Of those teams, only the Suns and Pacers were better at offensive rebounding. Larry Johnson is clearly the best player in Charlotte, leading the team in minutes, about tying Alonzo Mourning in offensive rebounds, having some of the most assists as a power forward in the NBA, and scoring 22 points per game on great efficiency. Him and the team were also decent in the playoffs, though obviously no match for the Knicks. What do you think about his and the team's season?

Larry Johnson was a nice player. He was at no point a top 5 player in the league, and was probably never more impactful than Alonzo Mourning.

[B-Mitch 30]

Larry Johnson was a nice player. He was at no point a top 5 player in the league, and was probably never more impactful than Alonzo Mourning.

I wasn't planning on having him as a top 5 player, and I think Mourning was more impactful, especially at their peaks, but I think Johnson was one of the best offensive players in the league this season, and better than Mourning in that regard (that turnover percentage).

[lessthanjake]

No, you keep trying to attribute nonsense arguments to him based on slight imprecisions in his choice of language — e.g. offhandedly saying “big jumps” must mean his stance is that by rule RAPM increases with increased sampling, and thus if he points out where several players saw declines with increased sampling, then he is now arguing against himself. Either he did not notice you doing that, or he is charitably letting it be.

There is no sincere reason to spend pages pushing incomplete “RAPM” with a predominant skew to one player/team.

I actually attributed a perfectly reasonable argument to him, and explained very clearly in multiple posts what argument I believed he was making and why I thought it was a very valid point but not a major factor in this particular instance. OhayoKD—who I’m certain we can both agree is not even remotely shy about correcting people about anything—responded to that multiple times without suggesting I had misunderstood anything, and only asserted he was actually making a different (and I think non-sensical) point after I pointed out that the information OhayoKD had pointed to actually supported what I was saying not vice versa. Only then was there some objection that actually a different argument was being made. And, to be clear, this isn’t even our first rodeo on this particular issue, and I’ve said the same things in the past about it and also never been told in those instances that I misunderstood anything. People can draw their own conclusions from that. In any event, it doesn’t really matter much, because whether it was the original argument or not, OhayoKD is free to make this other argument. I don’t think it makes much of any sense, though, nor has it been explained in any real way. I also think it’s just an odd thing to fight the substance of the Squared RAPM stuff a bunch here, when the other available RAPM and RAPM-like measures we have—the RAPM-approximation using quarter-by-quarter data, the complete RAPM in 1997, etc.—show us something very similar with regards to Jordan vs. Hakeem, without this issue OhayoKD is raising being at all at play. The RAPM and RAPM-like data picture for these players is all quite consistent, and it is highly favorable to Jordan.

[AEnigma]

But none of that applies to 1993. To the extent you are trying to litigate 1993 specifically, the Squared numbers basically just confirm, yes, Jordan was almost certainly one of the league’s most impactful players, and that was never disputed.

[OhayoKD]

This is a really curious post, since your argument is that RAPM goes substantially *up* as the sample gets larger than one year, and here you’re pointing to the fact that a bunch of peoples’ RAPM values went *down* as the sample got larger than one year. I think you got yourself confused here, and argued against your own point.

No? The argument was about what happens when a player's proportion of a total sample goes up. The more a player dominates a sample, a more they are pulled away from the mean.

The more years you play, generally your portion of the total sample shrinks. To whatever degree there is utility in using representatively distributed samples to model unrepresentatively distributed ones, what you would be looking for is drops.

I don’t think this is the point Bad Gatorade was making in what you quoted, nor does it make much of any sense to me (while the related point I take Bad Gatorade as making there that I addressed does actually make good sense, for reasons I’ve explained in this thread).

Nah, I'm pretty sure that is exactly the point being made here.

And yeah, the fact that it seems like this data is very Jordan centric in terms of game sampling means that there's more confidence in letting Jordan stray from the mean

They also said this shortly after...

brb, overrepresenting Chicago Bulls samples so that I can pull Jordan away from the mean

:/

I can ask them to clarify if u want but

And, in any event, in the pure RAPM I provided that you were commenting on, there’s not even any reason to think players’ proportion of the total sample would necessarily go up or down as the number of years goes up.

There is reason to think it for the top players generally

So your observations about the values in the RAPM source I provided really don’t map onto this argument you now claim to be making

You brought up RAPM to model/predict an upper-ceiling for the degree of potential inflation. I didn't realise you were trying to model it from "increases over everyone's sample getting larger", mostly because that doesn't make sense. Why would everyone getting more years sampled emulate the effect of one person dominating a sample?

[Djoker]

If the partial RAPM are that unreliable because they have oversampled Bulls games, then how come Scottie doesn't dominate the list and only Jordan does..? Pippen is 20th, 22nd, 2nd in RAPM in 1991, 1993 and 1996 respectively. His RAPM rankings are much worse than his actual standing in the league in two out of those three years. In the 1985-1996 RAPM where Pippen has a huge number of his games sampled, he's 22nd once again. lessthanjake also correctly pointed out that the Lakers actually have more sampled games than the Bulls in 1988 and yet Jordan is still ahead of Magic. Or that in 1991, the Lakers and Spurs have almost as many sampled games and yet both Magic and Robinson are well behind Jordan.

Not sure that this theory to discredit Squared2020's RAPM is all that convincing.

[OhayoKD]

If the partial RAPM are that unreliable because they have oversampled Bulls games, then how come Scottie doesn't dominate the list and only Jordan does..? Pippen is 20th, 22nd, 2nd in RAPM in 1991, 1993 and 1996 respectively.

Why are you only looking at Pippen?

The Bulls have 3 of the top 20 and 4 of the top 30 in 1991.

The Bulls have 3 of the top 30 and 4 of the top 50 in 1993

The Bulls have the top 2, 3 of the top 10, and 4 of the top 20 in 1996

lessthanjake also correctly pointed out that the Lakers actually have more sampled games than the Bulls in 1988 and yet Jordan is still ahead of Magic.

He does have a lead...which is much smaller than his general advantage.

Or that in 1991, the Lakers and Spurs have almost as many sampled games and yet both Magic and Robinson are well behind Jordan.

Not sure that this theory to discredit Squared2020's RAPM is all that convincing.

Not sure why 1991 Drob is worth mentioning here. Yeah, the advantage vs Magic there. No one is arguing Jordan can't end up looking like the best player in general via RAPM. The question is why we're deciding it matters for a down year in 1993 where 2nd place has half the possessions and almost everyone is working with a third.

You also never addressed this:

I love how apparently an entire season worth of games isn't nearly enough to stabilize RAPM but then at the same time WOWY off samples of anywhere between 5 games and 24 games are used as end all be all evidence that Lebron provides unsurpassed lift.

Not really sure where to start here.

1. Lebron's WOWY is not the end-all be all. He is also an era-outlier in RAPM despite spending several prime-years staggering with a similar teammate playing a position up, and also the most tested player situationally, and also an outlier in older and younger years, and also a Hakeem-level postseason riser. Jordan has a shot at one of those things. At the moment, Magic has most of them.

2. A season's worth of RAPM is not 82 games, it is a sample of what will overwhelmingly be 5-8 minute snippets of games stitched together with a few full games here and there. And whatever output comes from there is artificially curved towards 0 as an artificial scale is created to approximate adjustments to said 5-8 minute snippets. And in this case it is a sample dominated by one team that like most good teams of the era platooned a bunch. And in this season you are using it against a player who has 1/6th of the sample.

The main benefit of RAPM is it's stabler, but the trade-off of stability is you are no longer tracking real season to season impact, you are tracking a baseline. Even if you want to take that baseline at face-value, the 1-year outputs are artificially fitted to that baseline.

That actual stretches of 82 or near 82 games consistently have Jordan's teammates much better than the bits based on him missing a few minutes a game is a pretty big hint lineup effects are at play here.

I'm very confused where you got the idea that RAPM uses full games of off tbh. Or that the argument for Lebron as an era outlier is exclusively about WOWY