let's talk about small sample splits (again)

or, even better, mid-size sample splits that you would think would be big enough to draw conclusions from, but aren't

In the Dak-Zeke era, the two players have been involved in 1,520 plays on 1st and 10.

Zeke's plays (rushes and targets) account for 787 of those 1st and 10 calls. The remaining Dak plays (rushes, sacks, penalties, throws to non-Zeke targets, etc.) account for the other 733.

That's a lot of plays! What's more, they came in very comparable situations (all the exact same down and distance). So you'd think we could learn something new from, say, charting each of these series outcomes based on field position.

Before I post the chart, here are the averages:

Dallas Cowboys, 2016-2019

Series conversion rate when...

...starting with a Zeke play on 1st and 10—74.6%
...starting with any non-Zeke play by Dak—76.5%

These simple averages reflect a team that is moving the chains at a pretty good clip after both play types. They may have earned a first down on that same 1st and 10 play, or they may have converted later on the same series.

But not all field position is created equal. If you look at the NFL as a whole, you see teams convert on fewer series when they advance further into the red zone, for example. So for Dak and Zeke's plays on 1st and 10, we'd want to plot the team's conv. rates by field position.

Here's what our sample of 1,520 plays looks like—it's a total mess!

Note how the gray confidence intervals overlap. There's no part of the field where we can be sure which option was better, based on this specific, limited sample of series successes.