Earlier this offseason at Football Study Hall, Ian Boyd detailed some subtle changes coming to the TCU offense. With the addition of WR coach Malcom Kelly and the emergence of Taye Barber as a true #2 threat in spring ball, TCU looks to move their passing game downfield, incorporating some of the more dynamic and vertical offensive concepts the Big 12 has become accustomed to.
In fact, Gary Patterson himself has hinted at this new vertical offense: the coach has admitted that TCU is throwing the ball downfield as well as they ever had and touted the quality of competition at the quarterback spot. Whoever wins the starting job (my bet: Delton takes the first snap), it appears that TCU’s offense will be geared more towards deep threats than it’s been in the past.
But wait, TCU returns five starters on the offensive line and a three-headed monster backfield of upperclassmen; surely, the team whose rushing attack outperformed their passing attack, returning so much, would want to increase rushing volume?
Let’s take this idea head on, using 2018 play-by-play data and the nifty EPA data I introduced last week: Does TCU rush too much?
Each game, a football team implicitly faces a simple optimization problem. A team calls plays to maximize expected value subject to a couple simple constraints. For our purposes, the team is choosing a mix of run and pass plays to maximize expected value (EPA per play * share of plays), subject to the fact that the share of run and pass plays must equal one (that is, you must pick a either a run or pass every play). Does this oversimplify? Yes, but the simplification meaningfully isolates an important facet of decision-making.
Football and analytics nerds on Twitter will tell you that early down rush rates are indicative of a lacking passing game, and that teams too often rely on the rush. Let’s see how that bears out in college football:
The linear fit, the blue line, is downward sloping, indicating a negative relationship between early down rush rate and passing success rate: teams who rush on early downs more often are worse at passing. That’s probably more descriptive than causal, but still provides some guidance into proceeding towards answering our question: early down rush rate matters for the passing game.
TCU lies directly along the trend line above, just over 45% in rushing success and just under 45% in passing success, middle of the pack in both categories. On all downs, TCU’s offense ranked 86th in rushing efficiency, 113th in marginal explosiveness for rushing (101st, 71st for passing). TCU was more successful on the run than the pass last year, in terms of efficiency, but in terms of expected points added, the opposite is true.
I’ve filtered 2018 college football play-by-play data from CollegeFootballData.com, limiting observations to first and second downs, between the 20s, not in garbage time. That’s a severe restriction on an already limited sample, for sure, but it provides the clearest picture of team style by stripping away context that dictates what teams do. Of course, we don’t want TCU to pass more on third and short, or run more on third and long. Early downs and close games, out of the red zone/backed up zone generate the clearest picture of a team’s decision-making.
I constructed an “expected value” for a run and a pass for TCU. Simply put, it’s just the average EPA from a run times the rush rate plus an average EPA from a pass times the pass rate:
Expected Value = (EPA_rush * Rush %) + (EPA pass * Pass %)
TCU ran 453 plays in between the 20s, against FBS opponents, in non-garbage time this season. On average, the expected value of a play in this sample for TCU is (-.0146 * .42) + (.156 * .443) = -.006132 + .069108 = .063 expected points added, slightly positive.
Multiple approaches lend an answer to the question of whether TCU rushes too much. First, given the relative EPA of a rush and a pass (-0.146 for a rush, +0.156 for a pass) for TCU’s 2018 offense, the naive answer is: yes, TCU should pass more, because the expected value of a pass is higher than a rush. This is a naive approach, but it’s not a bad approach! There’s definitely an argument here.
Let’s go a little deeper, and incorporate success rates, because, after all, one might believe passing is a riskier enterprise.
Consider a dramatic increase in passes, say 15%, which corresponds to about 68 additional passes over the course of the season (17 a game) for TCU. Using this naive approach, TCU would replace 19.49 expected points added with 42.17 EPA, a net of +22.663 across the season. That’s a non-trivial swing!
We can make this analysis more robust by factoring in the mean loss of EPA on a rush and a pass, which will round out what might happen when we change strategy a little more fully. The formula we’ll use is:
EXPECTED VALUE = 68 * [EPA(success pass) * (sr_pass) + EPA(fail pass) * (1-sr_pass)] - [EPA(success rush) * (sr_rush) + EPA(fail rush) * (1-sr_rush)]
In words, we are subtracting the difference between the Expected Value of EPA on a pass and on a rush, and multiplying that out by 68 plays over the course of the season. Using this method, by rushing 15% more, TCU would’ve replaced -12.92 expected points added in the run game with +4.44 EPA in passes, a swing of almost 17 points!
Aw Heck, Man
The shrewd reader will be asking, about this point, “this is all good and well, but we’re taking success rate as a given; wouldn’t a change in run/pass tendencies affect the success rates of runs and passes?”
Here, I’ll employ perhaps the simplest application of the Heckman Selection model you’ll see in the wild. The Heckman model just addresses the idea of non-random selection; here, I’ll assume a “worst case scenario” replacement of runs with passes, assuming most of the rushes we replaced were good rushes, and the passes we replaced them with are bad passes.
Being generous, let’s say a 15% decrease in rushes, all of them “bad rushes”, would increase success rate on rushes to 45% (which is top 30 in the nation, so yes, this is an absurd generosity), and more passing decreases success rate to 40% (which would move TCU into the bottom third of teams nationally, again, absurd). Under this ridiculous circumstance, TCU would replace .3042 + -.45485 = -.15 * 68 = -10.244 expected EPA for rushing with .56 - .597 = -.037* 68 = -2.516 expected EPA for passing, which is still almost an 8 point swing!
Even in the worst case scenario for passing and the best case scenario for rushing, an increase in passing rates represents an absolute improvement in expected points outcomes!
Taking it to the Games
So, of course, a (conservatively-estimated) increase 17 expected points added doesn’t translate to 17 actual points spread out across the season, but it should translate to some positive increase in points across numerous situations. Looking at individual games can highlight specific points in the season where TCU was rushing too much.
First, let’s start with the two one-score losses: vs. Texas Tech and at Kansas.
In the Tech game, the expected value of a rush for TCU was -.340 EPA, and the Frogs rushed on 35.4% of plays in the sample, passing 64.6% of the time. You might think, “wow, that’s a lot of passes!” and you’d be right: 65.6% was TCU’s highest pass rate in a game in the sample. But 64.6% passing rate still wasn’t enough. The expected value of a pass for TCU against Tech was +0.0271, and by passing just 10 more times in the game, TCU could’ve improved it’s expected EPA by a net +3.67 points! The increase in expected points doesn’t mean TCU would’ve won the game or pulled 3.67 points out of the air to beat Tech, but rather means that in a one score game, in which the Frogs had the ball on the final possession, TCU left 3.67 points worth of expected value on the table by rushing too much on early downs.
Against Kansas, the expected value of a rush was -0.0473 points, and the expected value of a pass was +1.11 points (shout out Michael Collins and Kansas’s bad pass defense). Given rush and pass rates of 62.5% and 37.5%, respectively, increasing early down passes by just 10 would’ve increased expected EPA by 11.6 points! Again, that doesn’t directly translate to points, but it does translate to increased opportunities to score points against a terrible defense, and more opportunities to avoid the most embarrassing loss in college football.
Here are the other games where TCU “over-ran”, and could’ve strictly improved their expected outcomes by simply increasing their pass volume:
- Iowa State (+2.46 expected EPA from ten additional passes)
- Kansas State (+5.45)
- Ohio State (+7.66)
- Oklahoma (+4.66)
- West Virginia (+1.02)
Multiple methods of inquiry confirm the simple fact: TCU consistently left scoring opportunity on the field by rushing too much on early downs last season. An increase of 10 passes in TCU’s two one-score losses could swung their expected EPA by more than 3 points (vs. Tech) and more than 11 points (vs. Kansas).
As we enter the 2019 season, TCU’s offense remains a bit of a question mark. For TCU to return to its status as a contender in the Big 12, they’ll have to optimize talent and shake off their offensive woes. As the new downfield threat is rolled out, we will keep tabs on how rush rates change from last year, and with that, how TCU’s expected EPA benefits from offensive adjustments.