lorak wrote:I'm not doing it, because I'm not using RAPM at all. It was just response to your point about regression numbers to show that they aren't what you think they are. Really, how do you explain that Curry is better defensively according to regression? And why you ignored point about better team without Curry?
I think you need to be careful about telling the people who use the stat that they are doing it wrong given that you're not using it yourself.
On to Curry, there's a general conundrum here that's been recognized for as long as I've been on RealGM: If the presence of one player on offense means that the team can let the other players on the floor focus more on defense, then some of his impact will get listed as defensive impact by the +/- data even if his direct action is on offense. And the same is true in the other direction as well.
Now, the fact that that can exist doesn't mean that that's what's actually happening, but what it absolutely means is that you cannot take one side of the data at face value and then quibble about the other side. Just fundamentally, the fact that true impact can be in a way "misallocated" means that there is no basis for completely separating the two sides and thus concerns about what the data says on on side should immediately leave one asking about the other side.
You're trying to put the onus on me to explain X, as if I need to do that or else your argument wins by something like Occam's Razor, but the reality is that by far the simplest interpretation here is that the overall rating is the one with most credence.
Re: Ignored team better without Curry. But the team is also better with Curry and his overall rating is better too. Using the "better without" argument is thus not really a rebuttal. I wouldn't have brought up the overall rating if I didn't recognize that we have to factor in actual lift in the discussion.
I will note though: All things being equal, in general if two guys have the same rating, I tend to give the edge to the guy doing it on a better team. Not saying that that ends all analysis, but helping worse teams isn't the same thing as helping better teams.
lorak wrote: those who do value the stat might conclude that the guy with the better RAPM on the better team that also won the MVP rates higher than the guy you prefer.
I would like to see what explanation such people have for 1) different results from different RAPM models, 2) what difference is significant (for example 6 RAPM player is really "better" than 5.8?) and 3) that RAPM's value is it predictive power, not explanatory, so using it the way you do it (or I did in the past) doesn't make much sense.
Are these the metrics you're talking about?
Real Plus Minus
Evan Z's NPI RAPM
JE's Vanilla Multi-Year
If this is the case:
1) I'm uncertain about how precisely Real Plus Minus effects players like Curry and Paul, and this is why I've been very critical of JE ever since he moved toward xRAPM. It's just hard to see exactly what's going on. I think it's pretty clear cut that whatever box score, etc factor he includes it's helping Russell Westbrook LIKE CRAZY, but Curry & Paul are guys who look fantastic no matter what you do and I'm not sure if there is any particular bias between them. The fact that both Real Plus Minus and NPI RAPM favor Curry by a significant margin though is something real to me.
Again, not saying I base all my analysis on this data, but it's a mark in Curry's direction certainly.
2) The Vanilla Multi-Year analysis is what it says it is: Something that will favor a guy with more consistent track record in prior years more than the other two analysis. This is clearly the reason why LeBron has the lead going away from every one else: LeBron's been the king of this type of data. So yeah, I fully expect that this data is underrating Curry a good deal. The fact that he still manages to eek out a lead on Paul despite this is interesting too.
As far as what difference is significant, that's a good question and up for debate. Stating up front that I'm not really looking to use these numbers to say that X cannot possibly be more valuable than Y, if I look at the Real Plus Minus data, I'm likely to pretty well ignore difference much less than half a point. What's my basis for this? Just intuition based on experience. Could be wrong.
Re: "RAPM's value is in prediction not explanation." I don't buy this for a second. It sounds nice, but it only rings true based on a certain limited sense of the word "value" and to me the reason people like to say it is that it let's them avoid certain arguments.
Let's be clear, what bothers people about +/- stats is that they are quite clearly correlation rather than causation. Causation is what matters, not correlation, so why do we care about this data when we have actual causal data?
2 reasons:
1) Because in science realistically that's all we ever have. The goal of science is to try to infer causation from correlation. The fact that we can't ever truly know that we've done this correctly is the fundamental tenet of the modern philosophy of science, and it's why we start talking about things like sample size in the first place. What we're doing with +/- stats is not different from what we're doing in any kind of science where the data has noise in it, which at this point, is basically any form of cutting edge science.
So much of the resistance to +/- data to my mind simply comes from people not realizing that they are basically objecting to science in a manner very similar to those who reject evolution as "just a theory".
2) Because just because a player literally causes 2 points to be put on the board when he scores, that doesn't mean he's add 2 points that absolutely not have been there had he not been the one to take the shot. The box score tracks causal data, but the data itself cannot be mapped into actual causal impact in any kind of fool proof way, and since understanding causation is, as always, the goal, this means that once again we're stuck with correlation.
There's just no escape.
