Catching up on my reading and just got around to this piece by @PFF_Eric about whether good run-blocking OLs force defenses into more predictable coverages.

Wanted to make a point about something, though.

“Bet he’s going to say it’s selection bias...”

Hush, you, let me finish. https://twitter.com/pff_eric/status/1301497873343950849

So the most robust finding in the entire piece is that OLs that are good at run-blocking also tend to be good at pass-blocking.

“I bet it’s selection bias...”

Well yes, of *course* it’s selection bias. Harstad’s Razor: everything interesting is always selection bias.

BUT...

The interesting thing here is that the selection bias *cuts in the opposite direction of the observed effect*.

In other words, as strong as the observed relationship was, I’d expect selection bias is probably *diluting* the real extent of the relationship.

Consider a toy model: OLs have two jobs, pass blocking and run blocking, and to get hired in the NFL you need to hit some minimum threshold of goodness.

Let’s say both traits get graded on a 100-point scale and everyone with a combined score of over 150 makes the league.

Imagine a world with no relationship between pass blocking and run blocking.

You’d have some monster in the league who scored 100 in both. You’d have a run specialist who got 100 at run blocking and 50 at pass blocking. Or 50 at run blocking and 100 at pass blocking.

There’d also be guys who scored a 0 at pass blocking and a 0 at run blocking. (It me.)

But *those guys don’t make the league*. If you look just at guys who just barely make the NFL there is a perfect negative relationship between their pass blocking and run blocking.

Because of that perfect negative relationship at the selection threshold, the correlation for the total sample of NFL players would appear negative (despite already stipulating that in reality the correlation was zero).

This is Berkson’s Paradox.
https://en.m.wikipedia.org/wiki/Berkson%27s_paradox

Here’s a great illustration of Berkson’s. At most colleges, there’s a negative correlation between incoming students’ SAT scores and their GPAs.

Seems weird, right? Yeah, Harstad’s Razor: every weird finding is selection bias.

See?

Anyway, if Eric finds a moderately positive relationship between how good OLs are at pass blocking and run blocking, that probably means the real relationship is even stronger.

Because of selection bias, but in the opposite direction.

No larger point, I just like tweeting about Berkson’s Paradox.