I also think bad analytics has had a part to play in this. There are so many data guys, especially from an econ background, desperate to be the one to apply sabermetrics to football but failing to understand why it is doomed.

Compared to baseball, football seasons are far too short and football plays have far too many dimensions of freedom. Have only 16/17 games compared to 162 games a season means you have ⅒th of the sample data size to run regression models on and the likelihood of introducing systematic biases is huge.

Similarly, in baseball your single play is an individual pitcher versus an individual batter and there are much fewer degrees of freedom: the ball has to be pitched inside a small box, there are a very limited number of ways a non-foul hit can go, and a very small discreet number of outcomes.

In football, you have the entire offensive and defensive scheme to consider, with players moving anywhere in a 2D field, as well as consideration of separation, matchups, etc. Boiling all that down to a decision to run, pass, punt, kick and noting the field position, distance and down simplifies a ton of variables and who knows what that overlooks and oversimplifies?

With all that in mind, analytics in football is hard and the conclusions can be very oversimplified. It feels like the trend thanks to analytics and usage of stats like EPA etc has led coordinators to trend towards two big shifts:

1. Passing more often
2. Going for it on 4th down more often

The problem I feel is happening is that the entire football data analytics community is being blind to the fact that:

• Defensive schemes are not independent of offensive schemes. If the whole league starts passing more and being more aggressive on 4th down, the defenses will adapt and reduce the expected outcome of passing and going for it. But the analytics you used to make these assessments was trained on historical data and will completely miss these latest recent changing trends causing many teams to make the poor decision of passing too much when defenses have adapted. Eventually that data will start to factor into the regressions but on a time lag basis at which point the trends may have shifted again.

• The data used to train these models has survivor bias and leads to recommendations conditional on having a good (passing) offense. When the analytics says you should go for it or that passing in this situation has higher EPA, it is based on looking at the outcomes of pass v run / go for it v punt/kick in similar situations historically. Well that historical data only captures the plays that happened i.e. teams confident to go for it or pass on 1st down regularly. This biases the data simply because historically the data only exists conditional on that team having a stronger passing game.

In summary (TL;DR), the trend of using analytics has led more teams to pass more and be more aggressive on 4th down whilst overlooking these recommendations may not apply to their team (since the data was conditional on an era where defenses defended the run more and teams with a good QB). This may be causing teams that should pass it less, run it more, putting way more pressure on their QB than is necessary.