Play-by-play data allow us to develop statistics around what happens when individual players, or groups of players, are in the game. They allow us to look back at how certain lineups performed, and provide details as to why the performance was good, bad, or indifferent.
To give an example of the sort of analysis I am talking about, last season Texas scored nine points per 100 possessions more than its opponents when Isaiah Taylor was on the floor, and ten points per 100 possessions fewer than opponents when he was on the bench. (A large part of Texas' poor performance when Taylor sat: opponents shot 42 percent from three point range with the freshman point guard on the bench; this probably has little to do with Taylor's abilities.)
It turns out these numbers are endlessly fascinating. It gives us a different way of looking at the contributions of players. It is certainly not the last word in analysis (nothing is), but it is still interesting and may serve as a useful complement to other measures.
I was initially resistant to dive into this sort of analysis, for several reasons:
1. There are problems with the data.
2. Given problems with the data, complex correlation effects, and small sample sizes it is hard to know much we can safely infer from this information.
3. It seemed like a pain to do.
All three of those concerns are real. In general, most of the data you get from play-by-play logs are quite good, even if they possess a few idiosyncrasies and contain occasional mistakes. But deriving lineups from play-by-play logs can be problematic, and takes you into an entirely different and scary world. Players come in and out of a game quickly, often in large groups, and keeping up with all the changes can be hard for the person logging the data. My rough guess is that something like 5-10% of the lineups that you can derive from these data logs are wrong, at least with regards to one of the players listed. That seems like a lot of error.
But on the other hand, it also means that 90% or so of the lineups are right. By the standards of many of the things that get measured every day, a 10% error isn't such a big deal. Even more error-prone datasets can yield useful answers, provided we ask the right sorts of questions and don't fixate on small effects.
Then there is the question of inference. When we say that a particular team scores more points when a certain player is on the floor than it does when that player is on the bench, how much responsibility can we assign to this player? This is an important question, because the natural inclination with these sorts of data is to start assigning responsibility/blame.
I will leave this question unanswered, because I think it might be unanswerable. Rather than tackling this issue head on, I will dodge it somewhat. These numbers don't necessarily prove to us that Demarcus Holland should play a bunch, but they do perhaps provide some evidence that he should. (Much more on this below.)
Finally, concern number 3 was very real. The tools that I have generated to allow me to explore these data took a substantial amount of time to develop. It is built on top of a large data management infrastructure that has taken several years to develop to its current level of sophistication, and adds in some new data analysis tools that I have created to ease data exploration. The future looks very cool.
Those of you who felt that Javan Felix may have played too much or are in love with the game of Isaiah Taylor will find these numbers support your perspective. But those of you who think Rick Barnes gives too many minutes to Demarcus Holland will find that these data challenge your point of view.
The table below summarizes team results by Texas when each of the scholarship players were either in or out of the game.
|Player||Point Differential / 100 Poss When in Game||Point Differential / 100 Poss When on Bench||In Game - On Bench|
There is a lot of information to digest in that table. Rather than talking about every player on there (that could easily run to 10,000 words), I want to just highlight a few things that I found the most interesting.
Kendal Yancy and the strength of lineups where no one can shoot
Texas was strong when Kendal Yancy was in the game. The Longhorns scored 13 more points per 100 possessions during Yancy's 680 or so possessions than did opponents. The Texas offense was only two points per possession better with Yancy than without. The bigger effect is for the defense. When Yancy played, Texas opponents scored 95 points per 100 possessions, compared with 104 when he sat.
Quantitatively, the single biggest factor that improved the Texas defense when Yancy was in the game was opponent turnovers. When Yancy played, Texas opponents turned the ball over in 19 percent of their possessions. When he sat, the turnover rate was 15 percent. I estimate this difference as being worth about 4.5 points per 100 possessions on its own.
Some of this perhaps reflects the opponents Texas played when Yancy saw more minutes, some is luck, some is the effect of teammates, and some is likely due to the defense of the Texas freshman. It is hard to know how much credit to assign to each factor. But one thing we can say with certainty; when Yancy was in the game the Texas defense did well. We cannot say that it was because he was in the game, but it sure makes you wonder.
If we only looked at numbers, and didn't watch the games, we would have to stop there. But thankfully, we also have qualitative information to help us make inferences. Qualitatively, we know that Yancy is among Rick Barnes' best few perimeter defenders. So there is good reason to believe that Yancy plays a part in making the defense so effective when on the floor.
But there is another quantitative factor to consider. For about 60 percent of Yancy's possessions, Demarcus Holland was also on the floor. In the 403 possessions where these two players were defending together, the Texas defense was really good. Texas opponents scored 87 points per 100 possessions when both Holland and Yancy were in the game. Opponents turned the ball over in 21 percent of possessions, hit 39 percent of their two point field goals, and 25 percent from three when the two best Longhorn perimeter defenders were on the court.
So Yancy's numbers benefit from getting to play with Holland so much. For his part, Texas opponents scored exactly one point per possession during the 1745 possessions that Holland was in the game. So while Yancy sees the benefit of playing with Holland, Holland also benefits from playing with Yancy.
While we are on this subject, I think it is interesting to take a bit of a digression. The Yancy/Holland combo is great defensively, but it at least holds to potential to create trouble for the Texas offense. This is because neither Yancy or Holland was much help to the Texas offense, particularly when shooting from the perimeter.
But the thing is, Yancy/Holland lineups actually scored 110 points per 100 possessions. Only 19 percent of Texas shots were beyond the arc with both defensive stoppers in the game, but Texas made up for the lack of fire power by keeping down turnovers and being even better than usual on the offensive glass.
Taking this digression further, when Holland, Yancy, and Isaiah Taylor all played together, Texas was insanely good. In the 256 possessions that I captured with these three players on the floor, Texas scored 113 points per 100 possessions, while only allowing 77 points per 100 possessions.
I find this absolutely fascinating, and a bit counter-intuitive. It is hard to understand how you can have a functioning basketball team when not one of the three perimeter players on the floor can shoot. But this group didn't just function -- it thrived.
On offense, this trio basically refused to take three point shots. Less than 16 percent of Texas' shots were from long range when Taylor, Holland, and Yancy were all playing together. Instead, these lineups crashed the glass, took care of the ball, and went to the free throw line 0.54 times for every field goal attempt.
Meanwhile, the defense was suffocating. Opponents shot 14 percent from three point range, turned the ball over 21 percent of the time, and didn't get offensive rebounds very often.
Now, the tricky part. Does this mean that Texas would be best served to play Taylor, Holland, and Yancy together for heavy minutes next season? I think the most definitive answer that I can comfortably make is "maybe." I worry about predicting too much from information like this. We only have 256 possessions of data to go on (which is the equivalent of a little less than four games). That is enough data to make you wonder, but far from enough to make you sure. This might be a statistical fluke. On the other hand, it is probably a good idea if Rick Barnes wants to experiment with these sorts of lineups early next season.
Speaking of Demarcus Holland...
Fans like to complain about Demarcus Holland. I am not going to call out any of our writers or commenters here -- but yea, some of you fall into this category.
But through the lens these numbers Holland looks pretty good. Texas was 7 points per 100 possessions better than the opposition when Holland was in the game, and dead even with the enemy when he sat.
Tempo-free stats, as well as qualitative observations, provide a fairly nuanced view of Holland's offensive game. He is a little turnover prone and doesn't shoot the ball very well, offering little to the offense in half court situations. His best attribute on offense is that he is reasonably dangerous in transition, with a team high effective field goal percentage of 58 percent on shots attempted within the first ten seconds of possessions that started with a live ball change of possession.
When Holland was on the floor last season, the Longhorns scored about 1 point per 100 possessions more than when he sat. Let's just call that a wash. Holland doesn't have a big role in the offense, potentially messes up the floor spacing with his shooting problems, but at least partly compensates for doing a good job of finishing on the break. The lineup numbers above suggest that Holland probably wasn't the anchor on the Texas offense that he is sometimes made out to be.
On the defensive end things are different. With Holland in the game, Texas allowed 6 fewer points per 100 possessions than it did when he sat.
If you were basing decisions solely off of these numbers (and I certainly don't advise doing that), then it would lead you to the conclusion that Holland was Texas' third best perimeter player last season, and the high number of minutes he played were justified. But let's not go that far. A more responsible statement is this: it is hard to argue that Holland hurt Texas last season by playing so much, when the team was clearly better during the possessions he played than it was during the possessions he sat.
At the other end of the spectrum...
In the same corners of the internet where Rick Barnes took heat for playing Holland so much, he also was criticized for relying heavily on Javan Felix. And here the numbers seem to support the arguments made by the critics.
When Felix the Texas offense was decent, scoring 107 points per 100 possessions, compared with 106 points per 100 possessions when he sat. And the defense suffered, allowing 107 points per 100 possessions, compared with 92 points per 100 trips when the Texas sophomore was on the bench.
Javan Felix heavily shaped the Texas offense when on the floor, taking just under 30 percent of the team's shots when in the game. The Felix effect on the offense was at least partly responsible for a few things. Turnovers dropped, because Javan Felix is pretty careful with the ball. Meanwhile offensive rebounding rates went down, which perhaps has something to do with the fact that Felix often took jump shots, rather than attacking the basket and pulling weakside rebounders out of position. (This is the opposite of the Isaiah Taylor effect, who's presence on the floor increased offensive rebounding percentage.)
While the Horns were fine on offense when Felix played, Texas struggled on defense when the sophomore guard was in the game. Opponents shot 38 percent from three point range when Felix played. Some of this is dumb luck, for sure, but the qualitative observation is that Felix was often not quite as quick on close outs as some of his teammates. At a minimum, these data do not contradict that observation. Additionally, when Felix played, opponent turnover rates were fairly low.
How much blame can we really place on Felix for Texas' defensive problems when he was in the game? This is a difficult question. The most reasonable answer is, at least some of it. Another interesting way to look at this is to look more closely at one of Rick Barnes' most common starting lineups, the grouping of Jonathan Holmes, Cameron Ridley, Isaiah Taylor, Demarcus Holland, and Javan Felix. I have about 300 possessions for this lineup in my database. Texas scored 111 points per 100 possessions, and gave up 107 points per 100, for a differential of +4 point. For the 240 possessions where Holmes, Ridley, Taylor, and Holland played without Felix, Barnes' team scored 107 points per 100 trips and allowed only 85 points per 100 possessions. That is +22 points per 100 possessions.
I think these lineup-based numbers are useful, provided they are used carefully. As with any metric, they do not give a complete picture, and must be compared with other information.
But these numbers do at least provide clues. Where the clues are most interesting are where we don't have other good measures of performance. The most obvious place to apply them is with perimeter defense, where conventional basketball statistics are almost useless.
The cases looked at in this post hint at a pretty interesting idea. At least for the case of Texas during the 2013-2014 season, perimeter defense was a big factor in determining the success of a particular player grouping. It might have been the biggest factor.
Rick Barnes might just know what he is doing when he is giving all those minutes to Demarcus Holland.