Play Types and Their Varying Importance

facebooktwitterreddit

Feb 11, 2015; Minneapolis, MN, USA; Golden State Warriors guard Stephen Curry (30) drives to the basket past Minnesota Timberwolves forward Thaddeus Young (33) in the first half at Target Center. Mandatory Credit: Jesse Johnson-USA TODAY Sports

Thursday morning the NBA released Synergy Sports play type statistics that detail the end result of all NBA basketball possessions for the 2014-2015 season. Available for both the offensive and defensive side on the team and player level, these numbers are the result of possessions split into 11 different plays types—Pick & Roll Ball Hander, Pick & Roll Roll Man, Isolation, Cut, Handoff, Putback, Off Screen, Miscellaneous, Post Up, Spot Up, and Transition. Each play type represents the action that ended the possession, not necessarily what occurred during it, and the criteria for each play is briefly defined on this page here.

As Seth Partnow noted in an earlier article, these play type statistics are not a tell-all and should be digested with a considerably thick side of context. Beno Udrih has a higher points per isolation possession (1.39), than Kevin Durant (1.24), for example. The fact that the play type splits only reflect how the play ended tell us a very incomplete picture of the whole possession. It’s still better, however, than having no information about the types of plays that took place, and I am having quite a bit of fun looking through the relationships between play types and overall efficiency regardless. Did you know that Paul Millsap is the most efficient pick-and-roll ball-handler in the league this season?

In order to figure out exactly what these play type numbers can tell us, I took to the team level numbers to figure out what types of plays were the most and least efficient. Below is a table that details the points per possession average of each play type, sorted by efficiency.

Cut1.189
Transition1.104
Putbacks1.080
Spot Up0.976
P&R Roll Man0.971
Off Screen0.893
Hand Off0.863
Post Up0.857
Isolation0.850
P&R Ball Handler0.786
Miscellaneous0.482

Plays that end in cuts and transition plays are by far the most efficient plays. These numbers confirm our common sense notions on the importance of transition, and how much easier it is to score at the beginning of the shot clock, before defenders have acquired optimal position. Plays that end in cuts lead the way with regards to efficiency, and by a considerable margin. Cuts may be the most efficient play-ending events because cuts tend to lead to shots that are closer to the hoop and after less dribbles. Remember, these play types only represent how plays end.

Next, I looked to see what effect, if any, a team’s frequency of certain play types had on overall offensive and defensive efficiency. I tried to evaluate this in a few different ways, but found little evidence of a significant relationship between the frequencies with which teams perform or allow certain play types, and the resulting strength of their offensive and defensive efficiency. That is to say that NBA teams finish plays with varying levels of each play type, and this variance of play type frequency alone cannot reliably predict whether a team is good or bad on offense and defense.

The efficiencies of each team’s play types, however, most certainly can. Although offensive and defensive efficiency cannot be predicted reliably based on a team’s play type frequency dimensions, using the efficiencies of a team’s play type can not only predict a team’s efficiency, but it can tell us the importance of each play type’s efficiency. There are some play types which are much more important for a team to excel at then others.

To gauge a play type’s importance, I took the efficiencies of each team on all eleven play types on both the offensive and defensive side of the ball, and ran two regressions that predicted offensive and defensive efficiency. I normalized the play type efficiencies with z-score normalization, both to stabilize variance and so that the resulting regression coefficients of the significant variables could then be used as a means of ranking the importance of each play-type. There is some multicollinearity among the play type efficiencies, but nothing I consider bad enough to the point that it’s worth addressing and losing interpretability, so it’s better to just keep potential multicollinearity in mind. The results of each regression are detailed below.

Offensive EfficiencyRegression CoefficientSignificance LevelLeague AverageNon-normalized Variance
Cut OE0.2470.4981.18900.006318
Handoff OE0.2990.4020.86310.006100
Isolation OE1.0180.0420.84960.003401
Miscellaneous OE0.4680.1850.48190.006114
Putbacks OE0.5190.1831.08000.005275
Off Screen OE-0.5030.2380.89200.004309
Post Up OE0.0840.8000.85700.003406
P&R BH OE1.3820.0020.78600.003834
P&R RM OE0.7790.0620.97100.010393
Spot Up OE1.3240.0020.97600.004635
Transition OE0.9840.0061.10400.002946
Defensive EfficiencyRegression CoefficientSignificance LevelLeague AverageNon-normalized Variance
Cut DE0.0610.8611.1890.003103
Handoff DE0.4240.1560.8630.006122
Isolation DE0.0310.9320.8500.003222
Miscellaneous DE0.7350.0220.4820.004134
Putbacks DE0.2660.3341.0800.005447
Off Screen DE0.4520.2550.8930.003475
Post Up DE0.5130.0930.8570.002154
P&R BH DE0.3770.2250.7860.002763
P&R RM DE0.7670.0210.9710.004168
Spot Up DE1.2360.0150.9760.001885
Transition DE0.7840.0131.1040.003363

The r-squared for the offensive efficiency regression is 92.4%, and the r-squared for the defensive efficiency regression is 87.6%. These r-squared’s are very high, and as they should be: They represent efficiencies predicted by efficiencies. If anything, the r-squared of each regression are a little low, and this could be the result of missing plays, mislabeled plays, differences in data or something completely different altogether. (I used NBA data for the predictors, but Nylon Calculus’ non-estimated efficiencies for the overall efficiency targets, and this could be a cause of the lower-than-expected fit.)

Irregardless, the regressions each produce rather sensible regression coefficients that can be treated as an importance ranking because the predictors were normalized. Here is the offensive table again, sorted by importance:

Importance RankPlay TypeRegression CoefficientSignificance LevelLeague Average
1P&R BH OE1.3820.0020.786
2Spot Up OE1.3240.0020.976
3Isolation OE1.0180.0420.850
4Transition OE0.9840.0061.104
5P&R RM OE0.7790.0620.971
6Putbacks OE0.5190.1831.080
7Miscellaneous OE0.4680.1850.482
8Handoff OE0.2990.4020.863
9Cut OE0.2470.4981.189
10Post Up OE0.0840.8000.857
11Off Screen OE-0.5030.2380.893

Just because a play is the most efficient, does not mean it is the most important in determining overall offensive efficiency. Plays that end in cuts, for example, despite the fact that they are very efficient, tell us very little about how good a team is on offense. I would imagine that this is because most NBA players are capable of scoring and finishing out of cuts at a high efficiency, and that this ability is present on both bad and good offensive teams. Putbacks OE, Miscellaneous OE, Handoff OE, Cut OE, Post Up OE, and Off Screen OE were not considered to be significant in predicting offensive efficiency at all.

On the other side of the coin, plays finished by the Pick & Roll Ball Handler are very inefficient. The differences in this play type efficiency among the teams, however, are very important at determining how good a team’s offense is. Imagine the difference between Stephen Curry (0.99 in PPP as P&R Ball Handler), and Ricky Rubio (0.64 in PPP as P&R Ball Handler), and think about the difference in ability either of them has to finish the play right out of the pick and roll, and the influence that ability has on the other areas of a team’s offense. Defenders are in a lot more of a hurry to help on Curry than Rubio. This really speaks to the importance of above average off-dribble shot creators.

Importance RankPlay TypeRegression CoefficientSignificance LevelLeague Average
1Spot Up DE1.2360.0150.976
2Transition DE0.7840.0131.104
3P&R RM DE0.7670.0210.971
4Miscellaneous DE0.7350.0220.482
5Post Up DE0.5130.0930.857
6Off Screen DE0.4520.2550.893
7Handoff DE0.4240.1560.863
8P&R BH DE0.3770.2250.786
9Putbacks DE0.2660.3341.080
10Cut DE0.0610.8611.189
11Isolation DE0.0310.9320.850

Predicting defensive efficiency is somewhat similar, but different in its own right. Being able to defend the Pick & Roll Ball Handler is not considered important on defense (possibly related to this idea), but reducing efficiency of the Spot Up play still very much is. (Spot up plays not only include catch and shoot jump shots, but can also include drives to the rim as well.) Defending the Pick & Roll Roll Man and defending Transition is considered similarly as important on defense as on offense. As we learn more and more from analytics, we have grown to discover that good defense is more about preventing high efficiency events from happening than about reducing their actual efficiency, and much of that is apparent here in the differences between predicting good offense and predicting good defense. Isolation efficiency, which was significant in predicting offensive efficiency, is the least important play type efficiency to influence on defense. This is probably because if you’ve forced the offense into a low efficiency play like an isolation (0.85 PPP), you’ve already won, and the actual efficiency of the play isn’t as important to your overall defensive strength as forcing the play type is.

In terms of availability in the public setting, these numbers are literally days old and we still have much to learn from them. These numbers are only reflective of how plays end, and should thus be considered only in this light. Still, they provide further interesting insight into how basketball in the NBA works, and how teams can construct themselves and strategize to have the best competitive edge.