APBRmetrics

Thanks for the reply.

By the chart here the open 2 is significant better than the contested 3 in the very last second of the shot clock About equal frequencies of these shots taken. I doubt coaches and players can be trained specifically to take the open 2 in the final second. They may not really have a lot of choice about it that late. But by the numbers it would help.

One irritation of the 538 article/blog post/filler is characterizing the issue as “short or long but never midrange” - what is a theoretical straw man. Another is the underlying Goldsberry legerdemain of: threes might maybe be bad because... Dirk Nowitski!

I make the point here because these lovely graphs (unless I am misunderstanding the labels) compare open mid-range shots to contested threes and surprise, surprise, on average such substitutions appear to be very wise decisions (when feasible)! Who coulda guessed it?

But this is not really the relevant margin (nor in an important way the correct counterfactual). To begin, you want to compare apples to more distant apples. It would be interesting to see the plot for contested mid-range (and share of the total mid-range attempts that these are) rather than open. These gains presumably would be significantly larger than the 0.2 PPS shown for most of the shot clock.

Now that's much, much better: http://fivethirtyeight.com/features/the ... nba-steal/

schtevie wrote:Now that's much, much better: http://fivethirtyeight.com/features/the ... nba-steal/

Very interesting. Just curious if anybody has tried to duplicate the regression results. A steal being worth 9 points seems really high, but I have no reason to doubt it at this point. Also, and I think one of the commenters called this out, but is it possible to show how much of that nine points is due to offense vs. defense? Much in the same way you can do with +/-?

schtevie wrote:One irritation of the 538 article/blog post/filler is characterizing the issue as “short or long but never midrange” - what is a theoretical straw man. Another is the underlying Goldsberry legerdemain of: threes might maybe be bad because... Dirk Nowitski!

I make the point here because these lovely graphs (unless I am misunderstanding the labels) compare open mid-range shots to contested threes and surprise, surprise, on average such substitutions appear to be very wise decisions (when feasible)! Who coulda guessed it?

But this is not really the relevant margin (nor in an important way the correct counterfactual). To begin, you want to compare apples to more distant apples. It would be interesting to see the plot for contested mid-range (and share of the total mid-range attempts that these are) rather than open. These gains presumably would be significantly larger than the 0.2 PPS shown for most of the shot clock.

Those graphs should be in the article: http://blog.cacvantage.com/2014/02/the- ... fense.html
(my apologies on the axis being reversed or I suppose depending on how you prefer, maybe that way is better then the graph I posted in the thread)

boooeee wrote:A steal being worth 9 points seems really high, but I have no reason to doubt it at this point.

Seems way too high to me, other people have run regressions and not found results that high. He's frustratingly vague about what exactly he did in his analysis. What he describes is this: he did a "With-Or-Without-You" analysis, but at the game level (i.e. he found seasons where a player had missed at least 20 games -- or 15 games in an earlier analysis that he did) and compared the team's results with and without the player. I've got huge issues with sample size there -- how representative are those 15-20+ games, with regard to the quality of the opponent and the home vs road location? Also, as we've mentioned countless times here, who's replacing the player in the lineup? Is an injured Bill Walton being replaced by Swen Nater -- or LaRue Martin?

To get the value of steals, he's apparently running regressions looking at how the steals stats of the player in question correlate with his team's steal stats. And doing that for all of the WOWY players that he's got enough data for, to get an overall value of steals.

I've got big misgivings about the sample size and the resulting standard errors. Adjusted plus-minus has some inherent advantages over his technique (at least if I'm understanding what he did): larger sample size due to looking at individual events rather than individual games, and the multivariate regression technique takes into account the quality of the other 9 players on the floor, which AFAICT his technique does not do. But even with those advantages, as we all know APM is plagued with large standard errors. He's looking at about 25 years worth of data -- but in any one year there's only a relative handful of players who'll have enough games played and missed to qualify for the WOWY analysis, and those players' WOWY stats can cover at best 41 games.

It's a step above Berri's regressions which simplistically assigned the value of stats at the team level to the player who accumulated those stats in the boxscore. Morris is attempting to measure the marginal effect of a player's steals and other stats on his team's outcomes. But I don't buy those results.

The one good thing is that he's doing a quick rehabilitation of Ricky Rubio's portrayal in so-called advanced stats, by saying that Rubio actually helps the TWolves a lot despite his poor shooting, thanks to his steals. A needed antidote to Goldsberry's premature anointing of Rubio as the Worst Player in the NBA.

But the bigger problem is that this is the sort of stuff that gives analytics a bad name. Premature release of results based on models that are only half-developed, resulting in proclamations of results that are not believable. Thus causing the models not to be believed -- and all too often, the analytical efforts behind them not to be believed. If I was an outside observer, reading Goldsberry's analysis of Rubio and Morris's analysis of steals, I'd say that hoops analysts are off their rockers.

Oh, come on. No one or two observers represents an entire field. Try finding 2 economics "experts" at random, who also agree on much.
There's always going to be a market for results that are surprising, shocking, counter-intuitive. In the bazaar, shoppers pick up whatever catches their eye.

Clearly there's something amok in "one steal is worth 9 points", however they want to say it. Steals are down almost every year. Is it because they're so 'valuable' that teams don't allow them as much?

It may be an artifact of the variables he's regressing. Points are accompanied by missed shots. Assists and turnovers are related. But steals are not offset by 'missed steal attempts' in the boxscore.

"Drawing a charge" is another one without an offsetting number. It's not yet a boxscore stat, but some do track it. Yet what's the alternative result? Letting a guy score unimpeded? Letting him score and giving him a FT?

knarsu3 wrote:

schtevie wrote:But this is not really the relevant margin (nor in an important way the correct counterfactual). To begin, you want to compare apples to more distant apples. It would be interesting to see the plot for contested mid-range (and share of the total mid-range attempts that these are) rather than open. These gains presumably would be significantly larger than the 0.2 PPS shown for most of the shot clock.

Those graphs should be in the article: http://blog.cacvantage.com/2014/02/the- ... fense.html
(my apologies on the axis being reversed or I suppose depending on how you prefer, maybe that way is better then the graph I posted in the thread)

Knarsu3, thanks for the lead. First a correction. I misstated the results of the preceding graph, what shows a premium, for most of the shot clock, of "only" about 0.15 PPS for contested 3s over open mid-range shots. And now, per the relevant graph in the suggested article, the premium, over contested mid-range shots is about 0.2 PPS (until the last few seconds of the shot clock when it shrinks to approximately zero).

Regarding Mike T's dissatisfaction with the Rubio piece, perhaps I was guilty of a bit of soft bigotry. I was just happy that, by contrast to the previous one, a basically true story was being told, rather than thinking of the particular shortcomings of the regression. My initial reaction, however, had been that an overly-complicated story was being told. Establishing the value of Ricky Rubio, based upon the of argument of "skill scarcity" (getting steals in particular) can be told persuasively without recourse to regression. Probably even more so. Meh.

This may be a naive question, but when I saw this quotation in the article:

In fact, if you had to pick one statistic from the common box score to tell you as much as possible about whether a player helps or hurts his team, it isn’t how many points he scores. Nor how many rebounds he grabs. Nor how many assists he dishes out.

My first thought was: if this is true at the player level, shouldn't it be true at the team level?

Because as far as I can tell, it is not. I did a few scatter plots of various per possession stats to team efficiency for the 2013-14 season, and steals per possession had a very weak positive association with team efficiency. Weaker than blocks/possession, assists/possession, defensive rebounding %, turnover%, FTRate, OppFTRate. The only thing I found with a weaker association to team efficiency than steals was offensive rebounding%.

My bottom line question, then (and I apologize if this is stupid or I'm missing something) is: why does stealing the ball seem like such an insignificant indicator of team success if it's such a crucial indicator of player success?

Or to put it another way, if we imagine the league as a massive With-Or-Without-You experiment (which is Morris's method), why are the teams that are "without" players that produce a lot of steals not much worse on average?

Mike G wrote:Oh, come on. No one or two observers represents an entire field. Try finding 2 economics "experts" at random, who also agree on much.
There's always going to be a market for results that are surprising, shocking, counter-intuitive. In the bazaar, shoppers pick up whatever catches their eye.

Clearly there's something amok in "one steal is worth 9 points", however they want to say it. Steals are down almost every year. Is it because they're so 'valuable' that teams don't allow them as much?

It may be an artifact of the variables he's regressing. Points are accompanied by missed shots. Assists and turnovers are related. But steals are not offset by 'missed steal attempts' in the boxscore.

"Drawing a charge" is another one without an offsetting number. It's not yet a boxscore stat, but some do track it. Yet what's the alternative result? Letting a guy score unimpeded? Letting him score and giving him a FT?

What I find really weird is that the values of each stat seem to be inversely related to their frequency (I know blocks are a little less frequent than steals - but everything else fits the pattern). TOTALLY reminds me of my early day analytics (somewhere between 10 & 12 years old) when I decided every stat set had equal value (pts=reb=ast=st=bk=abs[missed shots]=abs[TOs]). So, in value per stat (due to frequency), you'd have a bk>st>to>ast>reb>missedshot>pt.

Are you telling me I was smarter at 12 than I am now?

Seriously though - my weights have for years weighed steals/bks/tos/asts more than other metrics I know - just obviously not nearly as much as this article seems to show their respective worth. In my ratings, a guy like TJ McConnell in college can be a top 100 player nationally despite averaging 7 to 8 ppg based on solid ast & st #'s, and lowerish TOs. Well, that and playing on a great team that gives up few points w/ a strong SoS.

The article is poorly explained and gives off a major scent of confirmation bias (not saying Morris did, just that it reads that way)

Neil's blog posts at basketball-reference are still my favorite. Innovative/creative/informative.

My favorites are still here at apbrMetrics. People actually discuss things they're doing!

My analogy for 538 so far is one of those sushi places with the spinning table

Dr Positivity wrote:My analogy for 538 so far is one of those sushi places with the spinning table

I think I understand the analogy, but I'm not familiar with spinning table sushi. I have seen several places where the sushi is on a conveyor belt and you grab what you want as it goes by -- sort of like a cafeteria in that you grab what you want except you're already sitting at your table. Or sort of like dim sum but with a conveyor belt replacing the person pushing a cart by your table. What are some names of the spinning table sushi places?


As for the quality of the articles so far on 538: I haven't yet come across one that I really like, but Neal P. has apparently written a good one about baseball forecasts. I haven't read it yet, but TangoTiger and Mitchell Lichtman both praised it on Tango's blog, and praise from them, MGL especially, typically is a meaningful endorsement.