RAPM request thread

[Nathan]

One additional thought on the age adjustment in 15-year RAPM: it assumes that all types of players follow the same aging curve. This is a different and more significant issue than the aforementioned problem with injured players like Derrick Rose. What if star players, on average, have steeper aging curves than roleplayers? What if frontcourt players peak earlier (or later) than backcourt players? If there is any such effect, then 15-year RAPM which assumes the same aging curve for all players will be biased for/against different sets of players. I would argue that this is a significant point in favor of using many years of either NPI APM or previous year informed RAPM for creating SPM, instead of using 15-year RAPM.

[Statman]

One additional thought on the age adjustment in 15-year RAPM: it assumes that all types of players follow the same aging curve. This is a different and more significant issue than the aforementioned problem with injured players like Derrick Rose. What if star players, on average, have steeper aging curves than roleplayers? What if frontcourt players peak earlier (or later) than backcourt players? If there is any such effect, then 15-year RAPM which assumes the same aging curve for all players will be biased for/against different sets of players. I would argue that this is a significant point in favor of using many years of either NPI APM or previous year informed RAPM for creating SPM, instead of using 15-year RAPM.

Higher usage guys - particularly wings/guards from my work, have much steeper "production" curves than low usage guys, particularly bigs. The higher usage guards/wings from a young age normally become even higher usage & more often efficient through their peak, along with some playing time increase. Lower usage bigs don't tend to produce all that much more from their young seasons (rebounds & block rates stay fairly flat or even drop early in the normal age curve) - they just tend to PLAY more as they hit their peak seasons.

I have no idea how you'd fit player "type" into this form of age curve analysis, since we are talking +/- & not individual production rates & rate types. You'd really start mixing apples & oranges in this case, could be messy.

[Nathan]

Indeed, I think this is a very rich problem with no easy solution. You'd have to have something like 538's player projector to make many-year RAPM properly. If I do get access to a large sample of single-year NPI APM, that's one problem I'd be sure to look at, basically running a regression where the dependent variable is change in plus/minus, rather than absolute plus/minus. If you run such a regression with age as the only independent variable you'd get the standard aging curve, but run it with additional independent variables like height, etc., and you'd get a clearer picture of how players develop.

JE, would it be possible to post more years of NPI APM? I would like to know soon if you would be willing to do this, in particular because I have a collaborator who's waiting to hear back from me about research plans and our options depend on what data we have to work with. Thanks, as always!

[nanawacin]

I have a question: is +6 NPI RAPM in year X equal to +6 NPI RAPM in year Y? If not, then why and how we should understand that stat if in one year +6 doesn't mean +6 per 100 possessions?

[J.E.]

I have a question: is +6 NPI RAPM in year X equal to +6 NPI RAPM in year Y?

It's supposed to be understood as "6 points better than the average player", and I suppose the quality of the average player can somewhat vary. In general, a +6 player with a team full of 0's, playing against all 0's should outscore the opponent by 6, no matter the year

[nanawacin]

Thx J.E. and I have another questions.

Code: Select all

PI   RAPM   RANK
2002   0,5   73
2003   0,9   82
2004   0,3   106

NPI   RAPM   RANK
2001   2,8   23
2002   3,1   15
2003   2,7   22
2004   0,4   119

MULTIYEAR   RAPM   RANK
2002   4,6   5
2003   4,9   7
2004   3,9   12

That's TMac's RAPM and rank among all players in given year (NPI and multiyear are from google docs you published at the end of last season; PI is from your old website). Why he is in top 25 first three years according to NPI, but ranks so worse in PI or so better in multiyear?

That's not just one case, as for example LeBron 2004 has -0.8 NPI RAPM (#236), 2005 1.5 (#60), but 2005 PI RAPM is 2.6 (#36) and 2005 multiyear 3.3 (#24).

Also:
- why RAPM values are so different in each of these 3 data sets?
- is multiyear RAPM the same as PI, just with better (bigger sample) prior?

[Jinxed]

Can we get an update for the end of this season? Pretty please with sugar on top. Single year and multi-year.

[J.E.]

Single https://docs.google.com/spreadsheets/d/ ... sp=sharing
Multi https://docs.google.com/spreadsheets/d/ ... sp=sharing (less weight to older seasons, 3 seasons total)

[Nathan]

Hi JE, I've noticed a potential flaw in your APM: it seems to use a zero prior that is independent of minutes played, and as a result it tends to overrate players with very few minutes and underrate players with very high minutes. This is clear in the following graph, which shows mean APM (averaged over ~100 players) vs minutes played:



According to this graph, players with ~500 minutes played have worse ratings, on average, than players with very few (<50) minutes played, which doesn't make sense because we would expect the relationship between minutes played and expected APM to be completely monotone. What I think is happening is low minutes players are in reality very bad on average (maybe -2 or so), but their APMs are dominated by the prior of ~-0.65, so on average they are overrated. Similarly, though less noticeably, high minutes players are very good on average (maybe +2 or so), but their APMs are dragged down by the ~-0.65 prior.

Would it be possible to make a version of APM with a minutes-dependent prior? Short of that, could you post APM with a much much weaker prior, so that the prior doesn't significantly affect any players' ratings? I realize this APM would result in low-minutes players likely having spookily high and low APM values, but it's easier to work with in some ways because it's relatively free of biases, and the uncertainties in the ratings are a simple function of minutes played.

Thanks as always.

EDIT: To be clear, I'd really like access to APM with a very weak prior, so that I can simply treat each player's APM rating as a random variate with mean equal to their true plus/minus value and standard deviation a function of minutes played. I'm writing a paper on statistical plus/minus, and I'm currently getting bogged down in the process of trying to transform the APM data to remove all minutes-related biases. Even if I manage to do so successfully, I'm worried that it'll water down the paper which I really want to keep focused on the interesting statistical plus/minus stuff. I would be very thankful for this, and of course credit you appropriately in my paper.

[Nathan]

Do you think you might get around to this soon, JE? I know you do all this on your own time, so I understand that if you don't think my approach is promising it doesn't make sense for you to spend your time working on something for me. But if that's the case, I'd like to know soon so that I can start trying to find an alternative source of APM. Of course if you do provide me with the APM I need I will be infinitely grateful, and I will be sure to give you appropriate credit for that substantial contribution in any publication that results!

EDIT: to be clear, what I want is APM like this: https://docs.google.com/spreadsheets/d/ ... sp=sharing
for as many individual seasons as is convenient.

[mzellman]

Hi!

I've been looking at the RPM Wins metric on ESPN (http://espn.go.com/nba/statistics/rpm/_/sort/WINS), and wondering if there's a corresponding stat for RAPM, and how far back that could be calculated. Would it be possible to get something like total Wins for each player since, like, 2001?

Also, would that include postseason play, or be regular season only?

[Crow]

Has it been presented in terms of season wins for RAPM? gotbuckets.com does it.

other sources for the archived data at shutupandjam.net or or here https://sites.google.com/site/rapmstats/

[mzellman]

This is precisely what I'm looking for. Thank you!

Do you know of anywhere that might have the 2015 calculations, or will be likely to have the 2016 calculations after the playoffs are over?

[Crow]

https://www.dropbox.com/sh/11181n4avq5w ... 5.txt?dl=0
from https://www.dropbox.com/sh/11181n4avq5w ... NNHEa?dl=0

https://docs.google.com/spreadsheets/d/ ... li=1#gid=0

discussed on first page or two on this thread

If not exactly what you want, ask JE.

[mzellman]

Yeah, I have those. They don't include the SWAgR number though, which is what I'm after.

I may just transition to RPM Wins from 2014 on, though...