G+Boost: Measuring What Happens After the Pass

by John Muller

What makes a passer good? 

Completed passes are better than incomplete, and more passes seem like they’re probably better than fewer, so you might start with the proposition that a player who attempts a lot of passes with a high success rate is a good passer. Darlington Nagbe got a lot of this kind of praise after the group stage of the pandemic tournament.

Of course, we know that not all passes are created equal. When the Crew’s social media team points out that Nagbe not only completed a lot of passes with a high success rate but did it in the opponent’s half, the idea is that those passes are better defended and harder to complete than if they’d been in his own half. A player who completes hard passes is presumably better than somebody who only completes easy ones.

We can measure pass difficulty more precisely with an expected passing model, which takes into account the pass’s location and angle as well as some additional information like whether it was a longball in open play. This helps put passing percentages in context by telling us, for example, that when Nagbe completed 43 of 43 passes against Atlanta United last week, an average player attempting similar passes would have completed 41. A player who beats the expected passing model is in one sense a better than average passer.

But whether a pass is complete isn’t the only interesting thing about it. The only reason we even care whether an attempt winds up with a teammate or an opponent, unless we’re Catalan aesthetes, is that the answer causes a tiny change in the balance of who’s likely to score next. A pass’s other qualities, such as whether it moves the ball away from your goal and closer to the opponent’s, can also change that balance. Sometimes that matters more than which team winds up with the ball.

American Soccer Analysis’s new goals added model measures how every action affects a team’s chances of scoring and conceding, and the model doesn’t care whether a pass is completed except to the extent it affects those chances. Goals added (g+) gives us a radically different picture of Darlington Nagbe. Even though he didn’t misplace a single pass against Atlanta, Nagbe’s passes were mostly sideways, slowing Columbus’s movement and hurting the team’s chances of outscoring the opponent over the next couple possessions. The goals added model rated his passing that night as below average for his position. In fact, across three group stage games, no midfielder in MLS had a better pass completion rate than Nagbe, but no midfielder had a worse passing g+.

Does that make Nagbe a bad passer? It might not be that simple.

Because it’s built on event data, which only records information about things that happen on the ball, the goals added model doesn’t know the movements of teammates and opponents. It draws inferences about a passer’s situation from what’s happened in the possession so far, like how long it’s been since a turnover and how fast the ball is moving upfield, but it can’t actually tell what kind of pressure the passer is under, what passing options are available, whether the ball finds a receiver in space, or if it’s delivered to the right foot with the right timing and weight. There are a lot of things that can separate a good completed pass from a bad one besides whether it travels forward.

One way to tell the difference between good and bad completed passes is to look at what happens next. Not every pass has to move the ball toward the opponent’s goal, but that’s ultimately where you want your possessions to go. The reason that a short sideways pass played to the front foot of a receiver in space is better than a short sideways pass played behind a guy with a defender breathing down his neck is that the first receiver is more likely to have time to pick out a line-breaking pass and the second is more likely to pass backward or turn the ball over. Those are outcomes that the goals added model can measure.

Against Atlanta, the Columbus players who received Nagbe’s 43 passes earned a total of -0.01 goals added on their next actions. That doesn’t sound great until you see that Nagbe’s midfield partner, Artur, completed 45 passes whose recipients lost -0.10 g+ on their next actions in the same game. If you only look at the value of each midfielder’s own passes in the Atlanta game, Artur looks better, with -0.04 passing g+ against average to Nagbe’s -0.07. But if you combine the value of their passes with what happened next, Nagbe comes out ahead.

A look at the video suggests why. Of the 39 Nagbe passes I found on a quick cut of the game, 56% were made under pressure, but only 31% of his receivers made their next action under pressure. When the receivers were under pressure, they either passed backward or turned the ball over every single time, but when they were in space their next action was positive more than half the time. By using his press resistant ball skills to draw defenders before releasing teammates into space, Nagbe’s sideways passing could help drive possession forward.

Screen Shot 2020-07-28 at 9.32.34 AM.png

To capture this skill at setting up what happens after the pass, ASA created a new metric called g+boost.*

Goals added boost looks at teammates’ g+ on the next action after receiving a pass from a particular player, like in our Nagbe example, but it goes a couple steps further. First, ASA’s resident modeler Matthias Kullowatz controlled for the average g+ of each receiver, so that LAFC’s Eduard Atuesta won’t have a higher g+boost than Nagbe just because he gets to pass to league MVP Carlos Vela while Nagbe’s stuck feeding an average player like Pedro Santos. Then Matthias applied something called a random effects model, which in his words “cuts through all the random BS and makes reasonable estimates for credible measurements of repeatable skill.”

The resulting metric is fairly stable from year to year, even when players move to a different team, suggesting that it’s picking up a real skill. Among players with more than 2000 complete passes, the career g+boost leader is Mark-Anthony Kaye, who helped lead LAFC’s midfield to unprecedented heights last year, followed by MLS legends like Clint Dempsey, Ignacio Piatti, and Michael Parkhurst. There are a few surprises on the leaderboard (Victor Ulloa?) but also one name that’s no surprise at all. Coming in 13th out of 163 qualifiers is a passer who’s never had a negative g+boost season, who’s helped his 8,795 pass recipients earn 2.5% more g+ than they otherwise would have, Ohio’s own Darlington Joephillip Nagbe.

Does that make him a good passer? Maybe not, since ideally you’d want players to play passes that add value and set their receivers up to do the same, but it’s a lot better than pass completion percentages for measuring what he’s good at.

image2.png

*This project started last year as xPass boost, using ASA’s expected passing model to determine whether receivers’ next passes were completed more often than expected, but it turned out that Toronto FC analyst Devin Pleuler had already done that back in 2014. The most interesting thing we found in our update was that the MLS midfielders Gregg Berhalter has selected for the USMNT all stand out for helping teammates’ expected passing scores.