We have just come out of the first year in the era of league-wide SportVU data, including a sliver of which has been made publicly available. It seems then like a good time to take a first stab at measuring what this fun new data means. I decided to do some analysis using stabilized adjusted plus minus data known as RAPM regressed against a combination of traditional box score metrics, shot location data and SportVU data1.
It is such a good idea, in fact, that I am hardly the first to do this. Kevin Hetrick in collaboration with the mysterious Talking Practice, did this kind of analysis focused on the ‘D’ side of RAPM for his site Got Buckets. They concluded that the Rim Protection metrics that have been the focus of Nylon’s own Seth Partnow indeed look like the metrics most likely to add value to a player’s defensive impact on the floor.
On the offensive side of RAPM, Nick Neuteufel has run a similar analysis including on a stat he called PEAR (Passing Efficiency Above Replacement), that was stat developed with ideas from Aqeel Phillips, a stat that divides assists, secondary assists and free throw assists by the number of passes.
Essentially with only one year of data any empirical analysis has to be classified as ‘exploratory’. These are early returns only, and this particular analysis only speaks to whether a stat will be useful in predicting overall player efficiency. For example, catch-and-shoot data or pull-up jump shooting data had limited predictive ability for a player’s impact as measured by RAPM2. However, that data is very helpful in looking at playing styles and possible team fit when putting together or scouting a roster.
In addition, data such as turnovers per touch, which I wrote about mid-season last year, may be an improvement as a stand alone statistic to quantify a player’s propensity to cough the ball over. In a multiple regression analysis, however, we are implicitly holding all of the player’s other contributions constant so there is not a distinct advantage over simply counting turn overs committed per 100 possessions on the court.
After running through a number of regression analyses, I don’t think there are any game-changers in the public SportVU data on the offensive side. However, there may be some marginal improvements in our understanding of efficiency, especially in terms of passing. There are a couple of new data points tracking passes made available to us in the public SportVU data: secondary assists (Hockey assists), free throw assists, total passes, and passes per touch.
The one derivation of passing data that seemed to add more to the analysis on a fairly consistent basis than simple assists per 100 possessions, was one very similar to Neauteufel‘s PEAR3. PEAR is built off of adding the points created by different kinds of assists together then dividing by the total passes the player used to rack up those assists.
In this analysis I primarily looked at passing efficiency as the number of assists, hockey assists or free throw assists per pass thrown rather than points to parse the data more for analysis purposes. In the analysis the effect of assists to passes efficiency was increased with the inclusion of passing frequency as measured by another SportVU data point, passes per 100 possessions. Of course, that makes sense in the same way that shooting efficiency is more useful in higher usage players.
On a team level, Passing Efficiency is associated with higher offensive efficiency, with an R^2 of ~ 0.17. One other note in interpreting Passing Efficiency is that it is even more closely associated with Pace on a team level than efficiency with an R^2 of ~ 0.23 last year. When I looked at touches per possession last year I found that more touches correlated to both longer possessions and less efficient offense, leaving some question of the direction of causation.
However, a look at the top passing efficiency players do not necessarily come from faster paced teams, as seen in the table below with the top twenty efficient passers. The table also has the number of passes the player recorded per 100 possessions on the court and a rating of their impact measured by one of the weighted least squared models I ran.4 Manu Ginobili was most efficient passer in terms of creating assists last year, creating assists for his teammates at over twice the average rate of players in the study. Rajon Rondo rated the 18th most efficient in creating assists, but the highest Passing Impact due to his 108 passes per 100 possessions.
Other expected names on the top list include Stephen Curry, Ricky Rubio, and John Wall. One interesting thing on the list is that it is less dominated by point guards than a pure assist counting list would be. To be sure, passing is a position related responsibility, the top three centers in passing efficiency were Zaza Pachulia, Marc Gasol and Joakim Noah, who were barely above average.
None of the other SportVU stats provided by the NBA publicly were consistently significant when tested against the RAPM scores. But, that’s not to say they aren’t useful, interesting or potentially expansive of our understanding of basketball efficiency or context.
- Data via NBA.com and Jerry Englemann’s Stats for the NBA ↩
- Catch and Shoot Rate (Total Catch and Shoot Attempts divided by total shots) had a slight negative correlation in some of the regressions once points scored and the number of shots or free throws taken were included. Indicating, possibly, a small discount in the efficiency of non-shot creators ↩
- Point Created by Assist fared slightly better than simple assists as well. Points Created by Assist is simply the total points created by any of the three types of assist tracked by SportVU, it also adds credit for assisted three point shots by adding the total points. Both tend to be more informative in the regressions as judged by the stepwise regressions and add marginally to the model’s overall explanatory power compared to assists per 100 possessions. In the case of PCA-100 the advantage simply comes from a better accounting of the action on the court compared to the vanilla assist counter. However, that measure is so well correlated to assists any difference is not meaningful. ↩
- In addition to the passing variables, this model included Points, FGA, FTA, O-Reb, and TOV (all per 100 possessions), as well minute per game and an interactive variable between rebounds and three point attempt rate. The passing coefficients were 1.915 per 10% increase in passing efficiency and .028 per pass in every one hundred possessions. ↩