(James Kenney/Associated Press)
In the summer, I wrote an article aimed at examining what determined running back success, based on an article from FiveThirtyEight. The FiveThirtyEight article stated that "96 percent of yards-per-carry totals are explained by the offense’s field position and the number of men the opponent has in the box." But what about the other four percent that is not explained by those factors? To find a better method for evaluate running backs, instead of using classic statistics like yards and yards per carry, my goal was to calculate how many more yards a running back gets than we would expect based on factors they cannot control.
First, using NFL play-by-play data dating back to 1999, courtesy of nflfastR, I created a linear regression model to predict how many yards a running play would gain based on down, yards to go, yard-line (from 0-100), and personnel. Since I did not have data on how many men were in the box on each play, I used a dummy variable for shotgun, which is equal to one if the offense is in shotgun and zero if not, as a proxy variable for personnel since defenses likely have fewer men in the box when the offense is in shotgun.
The linear regression model showed that down, yards to go, yardline, and shotgun all had a statistically significant effect on yards gained at the .001 significance level. The predicted yards range from about two yards to nine yards with most of the predictions around four or five yards (as seen in the histogram below), which means that the predictions could be off by plenty on a play-by-play basis, but it is significantly more accurate when calculating yards per carry on a season-by-season basis.
The graph below shows expected yards per carry (Expected YPC), based on the model, vs. actual yards per carry (YPC) for running backs with at least 65 carriers this season. The dashed lines represent the averages for each variable, so running backs in the upper left quadrant have an above average yards per carry and a below average expected yards per carry.
The objective of this research was not just to predict yards gained on a specific running play, but to use the predictive model to evaluate running backs more effectively. The model is supposed to predict yards per carry based on factors that running backs cannot control, so the difference in actual yards per carry and predicted yards per carry would be the yards per carry for which the running back is actually responsible.
To do this, predictions for every running play, with the exception of quarterback scrambles, were calculated using my linear regression model. Then, after grouping by rusher and calculating YPC and Expected YPC, they were subtracted, resulting in a statistic that I called Yards Per Carry Over Expected (YPC Over Expected = YPC – Expected YPC). The higher the YPC Over Expected, the more yards per carry for which the running back is actually responsible. The below chart shows the top-15 running backs by YPC Over Expected. (YPC Over Expected may not be exactly the same as YPC – Expected YPC due to rounding).
EPA/play is also a widely-used, effective statistic to measure running back efficiency, so I compared it to YPC Over Expected. I ran another linear regression model, which showed a statistically significant relationship between YPC Over Expected and EPA/play for running backs. Below is a graph showing the EPA/play and YPC Over Expected for running backs this season.
The best running backs are going to be in the upper right, which is where you’ll find players like Nick Chubb, Derrick Henry, and Aaron Jones, who are first, third, and fourth, respectively, in YPC Over Expected. Duke Johnson is last in the NFL in YPC Over Expected as well as EPA/play.
There are some other factors that running backs cannot control that I did not have data for when creating this model, such as offensive line performance or play-calling, but my model for YPC Over Expected seems to be an effective way to evaluate running backs.
Comments