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; here is Part 2:
My initial article explored how and why steals are underrated as a box score stat. Despite the constant focus on points per game in popular discussion of players, steals are actually a better indicator of value than points. Way better.
One of the most common responses to the article was the suggestion that we should consider the amount to which steals are a product of “gambling” on defense. In the comments, reader Kyle Pulek succinctly expressed the concern:
I think something that you have to consider is the “risk” of attempting a steal. In a way, steals are valuable but if you try and miss you end up out of position and increasing your opponent’s ability to score on that possession. If you racked up 5 steals a game, it wouldn’t be beneficial if the other 45 possessions that you played you swung and missed and ended up behind the ball.
I was a little surprised with the frequency and intensity of this objection, as I didn’t consider it a big issue. Indeed, I raised the same point in Footnote 7 (although I don’t blame anybody for not reading every footnote):
Steals come at a cost as well: By gambling on defense, you sometimes give up a better shot if you fail. But, all things considered, they are probably closer to being “free” than points.
The reason this didn’t concern me is that the part of the analysis that establishes the value of steals relative to other box score stats is completely oblivious to the costs and benefits of a particular stat; the analysis only cares about the corresponding increase or decrease in the team’s chances of winning.
Indirect “with or without you” analyses are meant to avoid the thorny and often intractable causal complexities that lie between a thing and its effect on the bottom line. Of course, this approach has its limitations. For example, we don’t know whether steals predict a player’s impact because steals are more important than other things, or because the type of player who tends to get steals just happens to be better at helping his team win games than a similarly situated player who doesn’t. But we do know that steals predict a player’s impact extremely well, so if we’re concerned with making empirical predictions, they’re something we should pay attention to.
This question usually comes from the opposite direction: When discussing the predictive value of steals with statsy-types, one of the most common responses I get is that it’s probably because steals are a defensive stat, and thus one of the only windows into a player’s defensive ability that box score stats provide. One of the more surprising side-findings in my analysis was that steals don’t seem to predict much about defense at all. If there is any contrarian element to my analysis, it is my argument that their value may stem mostly from their irreplaceability instead.
Moreover, though some heavy hitters apparently disagree, going for a steal doesn’t immediately strike me as a very bad gamble. A player’s reward is ending his opponent’s possession and getting an even more valuable than normal possession for his team, while his risk is possibly giving up a better shot. Even if some of those shots are layups, the overall difference in expected value of a failed steal attempt from no steal attempt is going to be much smaller than the value of the successful attempt vs. no attempt.
The one big thing we don’t know is how many attempts a player needs to get a steal. This “stealing efficiency” may vary considerably from player to player, just as efficiency-focused metrics for other box score stats do. But the observed predictive value of steals suggests that the ratio is not so dire. If anything, I could see it going in the opposite direction:
Although unlikely, it’s delightfully possible that the extra value provided by players who get a lot of steals is mostly just a result of their willingness to try to get them.
Finally, if you think that steals come with considerable risk, you still have to account for their predictive value somewhere, and I think your route to a plausible explanation is harder than mine: Perhaps they actually correlate to other (immeasurable) skills so strongly that it overwhelms their individually negative nature? It’s theoretically possible, but it’s not an easy case.