My article “The Hidden Value of the NBA Steal” sparked a lot of debate. I’m responding to several comments and questions in four parts. We posted Part 1 on Monday and Part 2 on Tuesday. Here is Part 3:
My original article argued that steals are the box score stat that tells you the most (all else being equal) about whether a player helps or hurts his team. But common response to this claim came in the form: “Wait, _____ gets lots of steals, and _____ stinks right?” (sometimes with a question mark at the end, sometimes an exclamation point).
The blank was most often filled by Monta Ellis (the Dallas Mavericks point guard with a long-standing reputation of being a ball hog) or the Philadelphia 76ers. Let’s address players and teams separately.
For players who we think stink but who get a lot of steals, there are a few possibilities:
- The player isn’t as bad as you think he is. It is not uncommon that players who seem bad (particularly those who put up ugly shots or who fare poorly in high-optic situations) tend to be much better than they appear. This is the sort of thing we turn to statistical analysis to (imperfectly) help us find.
- The player in question does stink, as is apparent from everything he does poorly other than getting steals. In this case, his steals may be valuable, but everything else overwhelms that value.
- The player in question does stink, and he stinks compared to other similarly situated players who get fewer steals. Nothing in our analysis precludes this. Just as statistical analysis shows that students who get high GPAs in high school do better in college on average than those who don’t, it doesn’t mean that no high school wiz will struggle in college, or vice versa.
I’ll refrain from opining on Ellis in particular. For great stuff on him, see Kirk Goldsberry’s “The Revenge of Monta Ellis,” as well as this response by Jeff Fogle. But I’ll grant that it’s possible that he’s a bad thief who attempts steals too often and hurts his team as a result (per above). But even if we assumed that’s true, it’s not evidence against the central finding. Indeed, Ellis is included in the data, so his impact on the historical record is accounted for.
The team question, however, is very different. It’s true: The 76ers lead the league this year in both steals and horribleness. Reader Chris Heitzig, who performed some of his own statistical analysis, emailed:
In my own regression equation, I used data from NBA.com for this season (up till today). In my regression, I found that the number of steals a team records is not a significant factor in determining winning percentage (p-value=0.66). And this makes intuitive sense too. With pick pockets like Rubio and Brewer, the Wolves (5th in steals) should be one of the best teams in the NBA. So too should the 76ers (1st). But both won’t make the playoffs. But teams like the Blazers (last) and the Pacers (25th), are playoff teams and have had great success so far this season. Of course we could expand this over a number of years, which would make the model more accurate. But it doesn’t look like steals have been much of a factor this season.
I’ve looked at team season data back to 1980, and there is a fairly solid positive correlation between team steals and winning, though it’s definitely “in the pack” with all the other stats (nerd version: correlation between steals per game and SRS is .136 over the period overall, though in a regression to SRS it has a normalized coefficient in line with points, rebounds, assists and blocks — all with similarly negligible p-values).
It’s worth noting that these types of correlations are not predictive. Testing the predictive importance of a team stat requires randomized in and out of sample regressions, but in this case, this is unnecessary: Even if steals have no predictive ability beyond their direct impact on games for teams, it has no impact on my conclusions with respect to players.
In fact, it may even bolster my case that the predictive value of steals stems largely from their irreplaceability: At the team level, every stat is 100 percent irreplaceable (if your team doesn’t get steals or assists, no other team is going to get them for you). So there is no immediate reason to expect steals to be any more predictive broadly than anything else with comparable immediate effects.