Monday, December 15, 2014

Clutch or not clutch, that is the stat question

There has always been debate on what clutch is, how it is measured, who performs best in these moments, the list goes on...
I'm definitely not going to settle the debate once and for all in one blog post, but wanted to share a few thoughts and ideas instead.

When we talk about clutch in the NBA, a few names immediately come to mind, Jordan, Bryant, Bird, Miller and Horry. Countless lists can be found with the simplest search, all as subjective as the next ("that was an incredible play in that game!").

But can we actually measure it, and rank players by it? has a whole section dedicated to clutch stats on its website. A great first step but amidst all the numbers it's hard to compare players.

SBNation also tackled the issue, clustering players into recipients, creators and scorers (not mutually exclusive) during clutch time (Is Kobe Bryant Actually Clutch? Looking At The NBA's Best Performers In Crunch Time). The article stresses the importance of efficiency by placing all performances in perspective using possessions per 48 minutes on the x-axis. Here are the results for the 2010-2011 season:

Efficiency is a trademark at SB Nation, and Kobe is their primary scapegoat given this viewpoint:

SBNation's perspective is interesting and allows swift comparisons across players, but I feel that it lacks some rigor and robustness around these numbers. How large are the sample sizes? Are the effects significant? Which players are shouldering the most pressure and confronting it head-on? The author underlines these issues himself:

But all of that said ... how reliable are these numbers? There's a school of thought that firmly believes that "clutch" is in the eye of the beholder. They contend that as fans, we see things that may not actually be there. We see Kobe hit a step-back 20-footer and credit his clutch ability, when perhaps we simply should have attributed it to the fact that he's amazing at basketball (in the 1st or 4th quarter).
There are rigorous methods of testing for statistical significance. Rather than dive into those, however, a glance at some yearly efficiency trends can be just as telling.

I also came across this very nice post, Measuring Clutch Play in the NBA, on the Inpredictable blog which offers an interesting and elegant alternative. In a nutshell, the idea is to look at how each player's actions impacted his team's probability of winning the game, referred to as Win Probability Added WPA. Made shots, rebounds, steals increase your team's probability, while missed shots and turnovers hurt it. Some adjustments are required to clean up the cumulative WPA for each player (essentially comparing the impact of the same play under normal circumstances), but it does at the end provide an intuitive metric that makes sense and allows quick comparisons.

I do however have some slight concerns with this metric. The first is that, unless I misread, the metric is cumulative, so that players with more minutes in the clutch have more opportunities to modify their team's WPA. The second is best illustrated with a small example: with a few seconds remaining, if a player makes a two-point shot with his team down by 2 or down by 1, it will make a huge impact on the WPA: in the first case they're tied, likely to go to overtime with 50/50% for each team to win the game, in the other case his team leads by 1 and have a good chance of winning the game. But is it fair to credit the player with very different WPA in both cases? What really matters is that, under tremendous pressure, the player made the shot.

This in turn leads to another question: what was the likelihood of that shot going in in the first place? How frequently does that player make that shot under normal circumstances without the game on the line? How frequently do other players make the shot? How much does clutch pressure reduce the average player's chance of making the shot, and was the player able to rise to the occasion and overcome the pressure?

According to Stephen Shea in his book Basketball Analytics, "90% of teams performed worse (in terms of shooting percentages) in the clutch than in non-clutch situations." Can this me modeled? How significant is the effect?

I will try to explore this path further, looking into statistical models that would offer some elements of response to these questions.

But looking at all hat has been said it seems the debate originates from the fact that "being clutch" is never well-defined. Suppose we could at any point during a game give a score from 0 to 100 as to how good a player is. Suppose player A is at 90 throughout non-clutch times, but drops to 80 in clutch situations. Whereas player B is at 60 in non-cutch situations, but steps his game up to 70 when the game is on the line. Which is clutchier? The one with highest absolute value, or the one stepping up his game and taking the pressure head-on. Answering this would already be a giant step in the right direction.

In the meantime, please enjoy this youtube compilation of clutch shots: