Why Mike “Analytics” McCarthy needs the Cowboys to run less and pass more

What worked and what didn’t work for the Cowboys last season and what it means for 2021.

One of his selling points when Mike McCarthy interviewed for the Cowboys head coaching gig was the promise of incorporating more analytics into his decision making. And while successfully rebranding himself as an analytics guy, he cleverly avoided explaining what exactly he meant by “analytics”. In some ways, that’s similar to a politician promising “change” or “justice” – catch-all phrases that everyone will interpret differently.

In 1964, Supreme Court Justice Potter Stewart, struggling to provide a definition for obscenity in Jacobellis v. Ohio (1964), put it simply: “I know it when I see it.” But for analytics, not even that low standard applies – because analytics can mean almost anything, there is almost no way to “see” it.

If you boiled down advanced football analytics to a few bullet points, the results would look something like this:

• NFL teams tend to pass the ball too little and run it too often
• NFL teams should go for more fourth-down conversions
• There is no correlation between running the ball early and winning games
• Offenses run too often on first down

This is not rocket science. And it’s not exactly a groundbreaking insight either, I wrote about these exact four points over 10 years ago.

Perhaps by choice, but more likely by coincidence, McCarthy’s coaching had already been at least “analytics-friendly” prior to his rebranding as an analytics guy. In his last five seasons with the Packers, McCarthy sported a 56.8% pass-play percentage, the second highest rate in the league. Over the same period, Jason Garrett (“Yeah, we don’t use those stats within the game, it’s just not something we think is best for our team”) led the Cowboys to a 47.2% pass-play percentage, the lowest value in the NFL over that five-year span.

In 2020, the McCarthy-enhanced Cowboys had a 58.3% pass-play percentage, the sixth highest value in the league.

Same thing for fourth-down conversion attempts: Between 2014 and 2018, the Packers had the fourth-most attempts in the league while the Cowboys ranked 19th, and last year the Cowboys were tied for the second-most conversion attempts.

Expected Points Added (EPA) is the metric of choice for most of the analytically inclined, and while it’s still a long way from being a mainstream stat in football, at least it’s readily available from sites like ProFootballReference.com, That’s laudable for a sport where a Super Bowl-winning coach can say things like “Stats are for losers. Final scores are for winners,” and goes unchallenged by beer-bellied sportswriters who still think it’s 1975.

Regardless, today we’re taking an EPA deep dive to understand what worked for the Cowboys offense in 2020 and what didn’t, and what it means for 2021.

What is EPA?

EPA is calculated by taking the expected point value (based on the league average results for that specific down, distance, and field position) before a play is run, and then subtracting it from the expected point value after a play is run.

The EPA for any given play is a value between -7 and +7. A positive value means the play contributed positively towards a score, a negative result means the play decreased the team’s odds of scoring.

EPA starts with the basic premise that not all yards are created equal. For example, a three-yard gain on 3rd-and-2 is much more valuable than a seven-yard gain on 3rd-and-8. Any measure of success must consider the down and distance situation, otherwise you’re talking fantasy football and not advanced analytics

ESPN explained the concept in a little more detail when they started using EPA back in 2012:

To make the concept more tangible, here are some examples:

• From your own 20-yard line, an 8-yard gain on third-and-10 is worth about minus-0.2 EPA because you don’t get a first down; the same 8 yards on third-and-7 is worth 1.4 EPA for converting a long third down and keeping the drive alive. EPA knows that not all yards are created equal.

• A turnover on first-and-10 at midfield that is taken back to your own 20 is worth minus-5.5 EPA; a Hail Mary interception at the end of the half from midfield is not nearly as penalizing. EPA knows that all turnovers aren’t created equal, as well.

• A 60-yard pass play down to the 1-yard line on third-and-10 is worth 5.7 EPA because it puts you right on the doorstep of scoring. The subsequent 1-yard rushing TD on first-and-goal is worth much less, even though that’s the play that actually gets you the six points. Think about which play is more valuable to the offense (not in terms of fantasy football).

EPA is not “Success Rate”

“Success rate” (gaining 40% of the necessary yardage on first down, 60% on second down and 100% on third and fourth downs) is a term that is being used with increasing frequency by NFL observers who think they are doing analytics.

But underlying success rate is the fallacy of a “manageable third down”, which is something Cowboys fans still traumatized by the Garrett era should be especially wary of.

This is not a question of opinion, but a question of hard, historical data (even if many people are increasingly confusing these two things these days). The data clearly shows that outside of 3rd-and-1 there are no manageable third downs. Third-down plays with more than two yards to go are low-percentage plays in the NFL. Again, this is not a new insight. Additionally, third-down performance is not replicable over time, so by focusing your play-calling on getting to third down, you’re betting on a handful of low-percentage plays to help you win the game.

So the next time you hear a coach blaming a loss on the lack of third-down conversions, feel free to tell him:

Using success rate leads teams to focus on maximizing success rather than maximizing the likelihood of scoring – which EPA focuses on.

For example, on 2nd-and-short, EPA suggests teams should be throwing down the field. But in many cases they do not, preferring instead to run the ball and thereby improve their success rate (by setting themselves up for the dreaded “manageable third down”) and not necessarily their likelihood of scoring.

Today we’re going to look at the 2020 EPA for the Cowboys’ offensive personnel, and we’ll work with EPA per play, starting with the ground game.

2020 Cowboys running game EPA

Including turnovers on running plays, the total EPA for the Cowboys’ ground game adds up to 19.61 on 425 runs (excluding five QB kneels). That’s an EPA of 0.05 per run, which means the Cowboys get a positive contribution from the running game, but just barely. Here’s how that breaks down by player:

Andy Dalton and Dak Prescott show up with strong EPA/run, and before 2020, I might have advocated for more designed runs by the quarterbacks; Prescott’s ankle injury has relegated that to a silly notion that I don’t plan to entertain anymore.

Lamb’s strong number suggests the Cowboys will improve their likelihood of scoring if they find a gimmick player to give more touches to this year. Of course, just having anybody run those gimmick plays may not be the solution: Amari Cooper (6 runs; -0.26 EPA/run), Cedric Wilson (3; -1.72), or Blake Bell (1; -2.0) did not prove very effective on similar runs.

But what’s up with Ezekiel Elliott and Tony Pollard? At first glance it doesn’t look like either players is contributing much in terms of points added, so how does that compute?

The simple answer is that EPA is about points, not yards.

A better answer is that in terms of EPA, the average combined value of runs for most of the field is close to zero. Here’s why: a first-down play needs at least four yards to be break even in terms of EPA, and in the NFL last year, 52.7% of runs (3,564/6,727) on 1st-and-10 gained less than four yards.

So if you’re in a situation where half your runs have a negative EPA and the other half have a positive EPA, your total running EPA should be at or around the zero mark. So the Cowboys aren’t doing that badly with a 0.05 overall EPA/play.

In Elliott’s case, his five fumbles last year (-15.82 EPA) are a key driver of his overall result. Excluding those five fumbles, his EPA/run would at least be slightly positive at 0.02.

But there’s another reason for Elliott’s low EPA, and that has to do with how the Cowboys used him. Here are his numbers split by down and distance (excluding fumbles):

Elliott delivered positive EPA in short yardage situations (EPA/run on 2nd/3rd/4th and less than 4 yards: 0.32), barely managed a positive number on 1st-and-10 (0.05) and on average decreased the Cowboys expected points every time he was asked to run with four or more yards to go (-0.13).

In a world before EPA, it was a widely held belief that there were some important benefits provided by the run game: helping out the defense by controlling the ball, tiring out defenses with ground-and-pound, and, of course, the holy grail of run-first truthers: setting up the pass.

All of which has been thoroughly debunked by the statistical community, but facts and belief systems never mix well, and so the run-first acolytes continue to fantasize about establishing the run.

This next table is the same table we used to look at Elliott’s EPA by down-and-distance, only this time we’re looking at the total passing game:

From an EPA point of view, the passing game delivers superior results in all situations, with the running game being good enough to at least be an alternative on second and third down with less than five yards to go. And that passing game includes all the sacks, interceptions, and incompletions the Cowboys had last year.

So if you need chunk yardage, you go with the passing game, if you just need a handful of yards, it’s okay to hand the ball to your backs. That, in a nutshell, is what today’s analytics are telling the NFL, but it’s not entirely clear who is listening.

Dallas Cowboys passing game EPA

Same exercise for the top eight receivers from last season:

The data here suggests three tiers of receivers.

Tier 1: The top three receivers all average slightly over 0.3 EPA per target. These are okay numbers, but they are not great. In 2019, the top three receivers all had a much better numbers: Amari Cooper (0.54 EPA/Play) Michael Gallup (0.51) and Randall Cobb (0.50) all had great numbers. With a healthy Prescott, we should expect similar numbers from the top three this year, and we’ll look at the QB impact on receiver performance in 2020 a little further down this post

Tier 2: Tight ends and backup receivers make up this second group. Slightly below the 0.3 of the top three WRs, and still a much better option than the running backs.

Tier 3: The running backs, to nobody’s surprise, are not a particularly efficient part of the passing game. Of course, it doesn’t help that Pollard and Elliott frequently had to help out with checkdown throws on broken plays.

With four different QBs throwing the ball last year, there’s bound to be some variance in the WR performance based on which QB they were playing with. The next table shows the EPA/Play for each of the top three WRs based on which QBs were throwing their way. I have excluded two of Prescott’s interceptions and four of Dalton’s from the data, as the small sample sizes in the table would be skewed significantly by those INTs.

The data here provides some expected and some unexpected insights:

• Prescott performed better with the top three WRs than Andy Dalton did, but perhaps not by quite the margin that might have been expected. Prescott did perform better with the remaining receivers (102 ATT; 0.35 EPA/Play) than Dalton (127; 0.15 EPA/Play) though.
• Interestingly, Dalton seems to have had a better rapport with Amari Cooper (0.66) than Prescott did (0.41). The downside of course being that Dalton was much less effective with CeeDee Lamb and Michael Gallup. But a healthy Prescott and a second-year Lamb? Watch out, NFL!
• Garrett Gilbert and Ben DiNucci both were significant downgrades over Dalton and Prescott, and even the quality trio of Cooper, Lamb, and Gallup couldn’t compensate for that.
• Ball distribution is also noticeable: Where Prescott seems to have favored Cooper over Lamb and especially Gallup, Dalton pushed the ball equally to all three wide receivers.

So where does all of this leave us heading into 2021?

Apart from the hyper-obvious “pass more, run less,” the data here suggests a few things the Cowboys might want to look at heading into the 2021 season:

• 11 personnel: If your top three wide receivers deliver the highest EPA/play, you should look to put 11 personnel (1 RB, 1 TE, 3 WR) on the field as much as possible. It was already the Cowboys’ default offensive formation, and its use has increased under McCarthy. In 2019, the Cowboys were in 11 personnel on 67% of their plays (ranked 11th in the NFL). That number increased to 71% in 2020 (ranked 7th), but the Bengals (76%), Steelers (75%), and Chiefs (73%) show that there is still room to grow.
• Play-action them to death: With $90-million Ezekiel Elliott in the backfield, everybody still expects the Cowboys to run on 1st-and-10. So instead, why not run play-action, a read-option, or an RPO play on every first down? The opposing defense will key on Elliott, opening up room for quick completions over the middle where – you guessed it – CeeDee Lamb will be waiting.
• Quality backup QBs: Go get one.
• Use a tackle to block: If you need a TE to help you block, use a tackle instead. If you need a TE to run a pass route, use a WR instead (Hint: you have enough of those).
• Don’t throw to your running backs. They may be “dangerous in space” but that space is never behind the line of scrimmage.

And, oh yeah, pass more, run less.