Emerging team strategy

[Mike G]

Ah, yes. Having the coach in the mix causes the players' rates to be adjusted.
A -0.5 rating is below the average player value, perhaps about median -- avg 6th or 7th man?

I'd still suppose an aging factor would enter in. Players are not as good in their final season as they are in their first, on average; so their average annual change will be negative.

[mystic]

Mystic raises the objection that Jeremias' ratings are suspect inasmuch as they fail to incorporate player-aging effects. True enough. But I think that this effect, taken into account, would serve to compress the range of results. Tom Thibodeau, table-topping rating, must be biased upwards for having inherited a very young team.

No, that is not how it works. Rose naturally improves after Del Negro left and Thibodeau comes in.

http://stats-for-the-nba.appspot.com/players/755.png

As you can see, Rose has -1.4 and -1.5 in prior informed RAPM in 2009 and 2010 respectively. In 2011 he goes up to 2.2 and in 2012 to 3.5. In the 10yr study Rose is assumed to be 0.4 in all included years. So, better than under Del Negro and worse under Thibodeau. Somehow at least part of the difference must be shifted to other included players/coaches in the 10yr study. It is completely reasonable to assume that Thibodeau's high value is in part just the natural improvement of Derrick Rose.

The apparent fact of the matter is that "head coaches" (coaching staffs, really) aren't that important in the grand scheme of things. Furthermore, on average, this collective input into the NBA production function appears to subtracts value.

It is not possible to draw such conclusions, because we are dealing with a biased sample. We could say that each team can select a head coach with rather similar success, but we can't say at all that the coaching staff is not important. The selection bias has to be taken into account. There is no comparison to situations without coaching, only with different preselected coaches.

And this (non-positive) result should be robust, given that it is based on 10 years' worth of data, no?

As I pointed out, there is a big issue with such long sample. Without a proper development curve for each player the results are basically becoming worthless. It is like superposition of waves with different wavelength. As long as you are clearly below the wavelength of the longer wave, you will not see much of an effect, but once you have a larger sample than the wavelength of the longer wave, you will see amplification and reduction.
In average players are about 6 yrs in the league. Assume a player needs 2yrs to develop then has 2 peak years and ends with 2 yrs of decline. Now, let assume such a player played his developing and declining years on a different team than his peak years. What will happen? The player will be assumed to be worse in his peak years than he really was, which will amplify the value of the players on the team he had his peak years with. On the other hand the players he played with during his developing or declining years will see a reduction. And even with players staying longer, we see such effects. Shawn Marion for example played his absolute prime years with Steve Nash, he declined after that and was weaker before. The 10yr study has Marion in average clearly weaker than what the pior informed values for him show during the Suns years with Nash. But somehow that must be compensated to keep the error low. It is reasonable to assume that Nash's value get a boost. Also, that Wade or Bryant getting their values amplified by a declining O'Neal.


And then again, what kind of value does a 10 yr study have anyway? Can I use it to explain results from the past, from specific years? Can I use it to make a prediction?

[Crow]

The 10 year results for coaches can provoke questions about why the coach got a very high or very low rating. In most cases one can construct a plausible explanation but the construct is one thing and may not be a full or accurate view.

There may be other more focused coaching impact study approaches that could yield better results. Regular season vs playoff performance. Coach to coach comparison of performance against top teams. One could also compare the RAPM results to the actual team win%s - expected / pyth. win%. And look at team factor level trends, especially on defense.

[schtevie]

To clear up some confusion:

Alas, some confusion remains. So, all the result of single year regressions posted have been shifted to find the best fit around zero, but the 10 year regression, what throws "coaches" into the mix, has not been.

Assuming this is correct, is there any inference based on past shifting that might suggest what the appropriate correction might be for the 10 year data? Can it be understood to be positive or negative, and is there a reliable approximation that could be provided, short of doing the actual work?

No, that is not how it works. Rose naturally improves after Del Negro left and Thibodeau comes in.

I don't think we disagree on the implication of the non-inclusion of aging curves on coaches ratings, do we? Those coaches in the 10 year sample who, on average, coached young teams (when players are improving "just" for having gotten more mature) will have their estimates biased upwards, and vice versa. On this account, Tom Thibodeau likely appears a bit better and Jeff Van Gundy, for example, a bit worse.

The apparent fact of the matter is that "head coaches" (coaching staffs, really) aren't that important in the grand scheme of things. Furthermore, on average, this collective input into the NBA production function appears to subtracts value.

It is not possible to draw such conclusions, because we are dealing with a biased sample. We could say that each team can select a head coach with rather similar success, but we can't say at all that the coaching staff is not important. The selection bias has to be taken into account. There is no comparison to situations without coaching, only with different preselected coaches.

Perhaps I wasn't clear, but I still think there might be disagreement on the point I was trying to make. I was not trying to make the point that coaching isn't required in NBA basketball. Of course it is. What the regression results quite clearly demonstrate, however, is that "best practice" coaching (as judged by conventional wisdom) doesn't seem to contribute all that much on the scoreboard: it's the players that matter (and by implication the GMs, e.g. Popovich, Gregg).

I have absolutely no doubt that Phil Jackson is a very good coach. (11th out of 99 over the past ten years, sez the results.) Surely excellent at managing over-sized egos. But the regression results say that over the course of his career, his (and his staff's) efforts have added about 0.8 points per game. That's not nothing (and relative to the average coach, if that's the comparison, it's a bit better still) but does that make him worth $10 million (or however much he commands) plus the extra few million for his assistants? More generally, are coaching expenditures, league-wide, rational? Are the benefits from the average (multimillion dollar) NBA coaching expense equal to the benefits gained from -0.5 points per 100 possessions (or whatever the correct, shifted value is)?

Finally,

And this (non-positive) result should be robust, given that it is based on 10 years' worth of data, no?

As I pointed out, there is a big issue with such long sample. Without a proper development curve for each player the results are basically becoming worthless.

I don't think this is right. The argument is not that the regressions based on the 10 year sample provide the best estimates of contemporary player values. The issue is whether the results for coaching can be understood to be in the right ballpark in terms of the average contributions of coaches. Yes, as already noted, the omitted variable of player maturation does effect the coaching estimates, but is there any reason to believe that this is economically meaningful and/or changes the general picture portrayed? This is not my sense, but it is an econometric point.

[J.E.]

Alas, some confusion remains. So, all the result of single year regressions posted have been shifted to find the best fit around zero, but the 10 year regression, what throws "coaches" into the mix, has not been.

That's not what I meant.
The regression itself does the shift to find the best fit. This happened for both the 10 year results and the single year results.
The single year results have been shifted _after the regression_ to avoid confusion.

None of this really matters though.
It's a tool to make predictions;
whether you predict players with values of (+1, +1, +1, +1, +1) to score X points against players with values of (-1, -1, -1, -1, -1)
or (0, 0, 0, 0, 0) against (-2, -2, -2, -2, -2), it's all the same.
It's just the difference that counts